Method and Apparatus for Determining the Frequency of Phrases in a Document Without Document Image Decoding

A portion of the disclosure of this patent document contains material that is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the U.S. Patent Trademark Office records, but otherwise reserves all copyright rights whatsoever.

The following concurrently filed and related U.S. applications having a common assignee are hereby cross referenced and incorporated by reference in their entirety.

"Method for Determining Boundaries of Words in Text" to Huttenlocher et al., U.S. patent application Ser. No. 07/794,392.

"Detecting Function Words Without Converting a Document to Character Codes" to Bloomberg et al., U.S. patent application Ser. No. 07/794,190.

"A Method of Deriving Wordshapes for Subsequent Comparison" to Huttenlocher et al., U.S. patent application Ser. No. 07/794,391.

"Method and Apparatus for Determining the Frequency of Words in a Document Without Document Image Decoding" to Cass et al., U.S. patent application Ser. No. 07/795,173 now U.S. Pat. No. 5,325,444.

"Optical Word Recognition by Examination of Word Shape" to Huttenlocher et al., U.S. patent application Ser. No. 07/796,119, Published European Application No. 0543592, published May 26, 1993.

"Method for Comparing Word Shapes" to Huttenlocher et al., U.S. patent application Ser. No. 07/795,169.

1. Field of the Invention

The invention relates to improvements in methods and apparatuses for document image processing, and more particularly to improvements for recognizing and determining the frequency of phrases in a document without first decoding the words or images or referring to an external lexical reference.

2. Background

In computer based electronic document processing, an attribute of the document(s) being processed which the operator often desires to know is the frequency with which some or all of the words occur. For example, Salton & McGill, Introduction to Modern Information Retrieval, Chapter 2, pp. 30, 36, McGraw-Hill, Inc., 1983, indicates that in information retrieval contexts, the frequency of use of a given term may correlate with the importance of that term relative to the information content of the document. Word frequency information can thus be useful for automatic document summarization and/or annotation. Word frequency information can also be used in locating, indexing, filing, sorting, or retrieving documents.

Another use for knowledge of word frequency is in text editing. For example, one text processing device has been proposed for preventing the frequent use of the same words in a text by categorizing and displaying frequently occurring words of the document. A list of selected words and the number of occurrences of each word is formulated for a given text location in a portion of the text, and the designated word and its location is displayed on a CRT.

An extension of this thesis is that knowledge of the frequency of sequences of words in reading order in a document, i.e., phrases, also is useful, for example, for automatic document summarization. Phrase frequency information can also be used in locating, indexing, filing, sorting, or retrieving documents.

Heretofore, though, word frequency determinations have been performed on electronic texts in which the contents have been converted to a machine readable form, such as by decoding using some form of optical character recognition (OCR) in which bit mapped word unit images, or in some cases a number of characters within the word unit images, are deciphered and converted to coded representations of the images based on reference to an external character library. The decoded words or character strings are then compared with dictionary terms in an associated lexicon. Disadvantages of such optical character recognition techniques are that the intermediate optical character recognition step introduces a greater possibility of computational error and requires substantial time for processing, slowing the overall word unit identification process.

REFERENCES

European Patent Application No. 0-402-064 to Sakai et al. describes a text processing device in a computer system for counting the occurrence of words in a text and displaying a list of repetitive words on a CRT. The list includes the selected words together with their number of occurrences and their locations in the text. In a case where word repetition is undesirable, an operator may substitute synonyms or otherwise alter the text by using search, display, and editing actions.

European Patent Application No. 0-364-179 to Hawley describes a method and apparatus for extracting key words from text stored in a machine-readable format. The frequency of occurrence of each word in a file, as compared to the frequency of occurrence of other words in the file, is calculated. If the calculated frequency exceeds by a predetermined threshold the frequency of occurrence of that same word in a reference domain appropriate to the file, then the word is selected as a key word for that file.

European Patent Application No. 0-364-180 to Hawley describes a method and apparatus for automatically indexing and retrieving files in a large computer file system. Key words are automatically extracted from files to be indexed and used as the entries in an index file. Each file having one of the index entries as a key word is associated in the index with that key word. If a file is to be retrieved, and its content, but not its name or location, is known, its key words are entered and its identifying information will be displayed (along with that of other files having that key word), facilitating its retrieval.

Concurrently filed U.S. patent application Ser. No. 07/795,173, now U.S. Pat. No. 5,325,444 to Cass et al., and entitled "Method and Apparatus for Determining the Frequency of Words in a Document Without Document Image Decoding," which application is incorporated herein by reference, describes methods and apparatus for determining word frequency in an undecoded document image based on segmentation of the document image into image units and comparing image characteristics of selected image units with image characteristics of other selected image units to determine equivalence classes of image units. The invention described herein extends this image based word frequency methodology to determination of phrase frequencies without document image decoding.

Accordingly, an object of the invention is to provide methods and apparatus for determining the frequency of occurrence of similar phrases in an undecoded document image without document image decoding.

It is another object of the invention to provide methods and apparatus of the type described that reduces or eliminates the possible introduction of unnecessary errors due to intermediate interpretation processes, and is more efficient than optical character recognition algorithms or bitmap correlation methods.

It is another object of the invention to provide methods and apparatus of the type described that may be performed in a short time, and are parallelizable.

In accordance with one aspect of the invention, a method for determining a frequency of occurrence of phrases in an undecoded document image is presented. According to the method, the document image is input and segmented into image units having undecoded information content. At least one significant morphological structural image characteristic for each of selected ones of the image units is determined, and equivalence classes of the selected image units are identified by clustering image units with similar morphological image characteristics. The sequences of equivalence class labels corresponding to all possible sequences of the selected units arranged in the order in which the selected image units appear in the document image are determined; and the equivalence class label sequences are then evaluated to identify the numbers of recurring subsequences, and thereby determine the frequency of each equivalence class label sequence.

In accordance with another aspect of the invention, the image units may be word units in a textual document, and a word unit is preferably evaluated by deriving a word shape representation of the word unit, which is either at least one, one-dimensional signal characterizing the shape of the word unit; or an image function defining a boundary enclosing the word unit, which image function has been augmented so that an edge function representing edges of the character string detected within the boundary is defined over its entire domain by a single independent variable within the closed boundary, without individually detecting and/or identifying the character or characters making up the word unit.

The equivalence classes preferably are determined by comparing selected morphological characteristics or combinations of characteristics, or the derived representations of the image unit shapes, with each other. The morphological characteristics can include image unit length, width, font, typeface, typeface cross-section, number of ascenders, number of descenders, or the like. In a particularly efficient comparison according to the invention, only the height and width dimensions of the aforesaid word shape representations are compared. The image units in each equivalence class are linked together, and mapped to enable the frequency of each to be determined.

In accordance with still another aspect of the invention, an apparatus for processing an undecoded scanned document image to determine the frequency of phrases in the document image is presented. The apparatus includes phrase frequency determining means for computing frequencies of phrases by utilizing non-content based word unit morphological image characteristics, and an output device. The word and phrase frequency determining means may be a programmed digital computer.

The present invention provides methods and apparatus for determining the frequency of phrases in a document directly from a stored image of the document, without first converting the document to optical character codes. The invention utilizes non-content image recognition techniques that compare morphological image properties of word image units without the need to correlate bitmaps.

The invention is not limited to systems utilizing document scanning. Rather, other systems such as a bitmap workstation (i.e., a workstation with a bitmap display) or a system using both bitmapping and scanning would work equally well for the implementation of the methods and apparatus described herein.

The major advances of the claimed invention include being parallelizable, enabling performance in a short time, and providing for extracting word sequence frequency information without first converting a document to character codes, such as ASCII codes. Also, the invention eliminates the possible introduction of errors due to intermediate interpretation processes, such as optical character recognition. The method is more time and cost efficient than OCR algorithms or methods based on bitmap correlation.

These and other objects, features and advantages of the invention will be apparent to those skilled in the art from the following detailed description, when read in conjunction with the accompanying drawings and appended claims.

The invention is illustrated in the accompanying drawings, in which:

FIG. 1 shows a flow chart of a method according to a preferred embodiment of the invention for determining phrase frequencies in text of a document without first converting the text of the document to character codes.

FIG. 2 shows an apparatus according to a preferred embodiment of the invention for determining phrase frequencies in a text of a document without first decoding the image units or text or converting the image units or text in the document to character codes.

FIG. 3A shows a generalized system diagram of an image processing system in which the present invention would find use;

FIG. 3B shows a block system diagram of the arrangement of system components forming one embodiment of the inventive word shape recognition system;

FIG. 4 shows an image sample of example text over which the inventive process will be demonstrated;

FIG. 5 is a copy of a scanned image of the example text;

FIGS. 6A, 6B and 6C graphically illustrate the process used to determine the angle at which the example text is oriented in the image sample prior for further processing, while FIG. 6D shows graphs of the responses taken from the example text, which are used to determine the angle at which the example text is oriented in the image sample prior to further processing;

FIGS. 7A and 7B respectively show the derivation and use of a graph examining the sample image of the example text to determine baselines of text within the image;

FIGS. 8A and 8B are flowcharts illustrating the procedures executed to determine the baselines shown in FIG. 7A;

FIG. 9 shows the scanned image of the example text with baselines indicated thereon after derivation from the data shown in FIGS. 7A and 7B;

FIG. 10 is a flowchart illustrating the steps used in the application of a median filter to the image of FIG. 4;

FIG. 11 is an enlarged pictorial representation of a portion of the image of FIG. 4, illustrating the application of the median filter;

FIG. 12 demonstrates the resulting image after application of a median filter, a process known herein as blobifying, to the scanned image of the example text, which tends to render character strings as a single set of connected pixels;

FIG. 13 shows a subsequent step in the process, in which lines of white pixels are added to the blurred image to clearly delineate a line of character strings from adjacent lines of character strings;

FIG. 14 is a flowchart illustrating the steps required to add the white lines of FIG. 13;

FIGS. 15A and 15B are flowcharts representing the procedure which is followed to segment the image data in accordance with the blurred image of FIG. 12;

FIG. 16 shows the sample text with bounding boxes placed around each word group in a manner which uniquely identifies a subset of image pixels containing each character string;

FIGS. 17A and 17B illustrate derivation of a single independent value signal, using the example word "from", which appears in the sample image of example text;

FIG. 18 illustrates the resulting contours formed by the derivation process illustrated in FIGS. 17A, B;

FIG. 19 illustrates the steps associated with deriving the word shape signals;

FIGS. 20A, 20B, 20C and 20D illustrate derivation of a single independent value signal, using the example word "from";

FIGS. 21A, 21B, 21C and 21D illustrate derivation of a single independent value signal, using the example word "red", which does not appear in the sample image of example text;

FIG. 22 shows a simple comparison of the signals derived for the words "red" and "from" using a signal normalization method;

FIGS. 23A, 23B, and 23C illustrate the details of the discrepancy in font height, and the method for normalization of such discrepancies;

FIG. 24 is a flowchart detailing the steps used for one method of determining the relative difference between word shape contours;

FIG. 25 is a flowchart detailing the steps of a second method for determining the relative difference between word shape contours; and

FIGS. 26A and 26B are respective illustrations of the relationship between the relative difference values calculated and stored in an array, for both a non-slope-constrained and a slope-constrained comparison.

FIG. 27 is a flow chart of a preferred embodiment of a method according to the invention for detecting function words in a scanned document image without first converting the document image to character codes;

FIG. 28 is a block diagram of a preferred embodiment of an apparatus according to the invention for detecting function words in a scanned document image without first converting the document image to character codes;

FIGS. 29A-1-29C-2 show three sets of character ascender structuring elements where: FIGS. 29A-1 and 29A-2 show a set of character ascender structuring elements of height 3 and length 5 where FIG. 29A-1 shows a structuring element with ON pixels along the bottom row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 29A-2 shows a structuring element with ON pixels along the bottom row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels; FIGS. 29B-1 and 29B-2 show a set of character ascender structuring elements of height 4 and length 5 where FIG. 29B-1 shows a structuring element with ON pixels along the bottom row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 29B-2 shows a structuring element with ON pixels along the bottom row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels; FIGS. 29C-1 and 29C-2 show a set of character ascender structuring elements of height 5 and length 5 where FIG. 29C-1 shows a structuring element with ON pixels along the bottom row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 29C-2 shows a structuring element with ON pixels along the bottom row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels.

FIGS. 30A-1-30C-2 show three sets of character descender structuring elements where: FIGS. 30A-1 and 30A-2 show a set of character descender structuring elements of height 3 and length 5 where FIG. 30A-1 shows a structuring element with ON pixels along the top row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 30A-2 shows a structuring element with ON pixels along the top row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels; FIGS. 30B-1 and 30B-2 show a set of character descender structuring elements of height 4 and length 5 where FIG. 30B-1 shows a structuring element with ON pixels along the top row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 30B-2 shows a structuring element with ON pixels along the top row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels; and FIGS. 30C-1 and 30C-2 show a set of character descender structuring elements of height 5 and length 5 where FIG. 30C-1 shows a structuring element with ON pixels along the top row and along the rightmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels, and FIG. 30C-2 shows a structuring element with ON pixels along the top row and along the leftmost column and one or more OFF pixels in a remaining location preferably separated from the ON pixels; and

FIG. 31 shows a horizontal structuring element of length 5.

The Appendix contains source code listings for a series of image manipulation and signal processing routines which have been implemented to demonstrate the functionality of the present invention. Included in the Appendix are four sections which are organized as follows:

Section A, beginning at page 1, comprises the declarative or "include" files which are commonly shared among the functional code modules;

Section B, beginning at page 26, includes the listings for a series of library type functions used for management of the images, error reporting, argument parsing, etc.;

Section C, beginning at page 42, comprises numerous variations of the word shape comparison code, and further includes code illustrating alternative comparison techniques than those specifically cited in the following description;

Section D, beginning at page 145, comprises various functions for the word shape extraction operations that are further described in the following description.

A preferred embodiment of the method of the invention is illustrated in the flow chart of FIG. 1, and apparatus for performing the method of FIG. 1 is shown in FIG. 2. For the sake of clarity, the invention will be described with reference to the processing of a single document. However, it will be appreciated that the invention is applicable to the processing of a corpus of documents containing a plurality of documents.

With reference first to FIG. 2, the method is performed on an electronic image of an original document 5, which may include lines of text 7, titles, drawings, FIGS. 8, or the like, contained in one or more sheets or pages of paper 10 or other tangible form. The electronic document image to be processed is created in any conventional manner, for example, by an input means, such as an optical scanner 12 and sensor 13 as shown, a copier machine scanner, a Braille reading machine scanner, a bitmap workstation, an electronic beam scanner or the like. Such means are well known in the art, and thus are not described in detail herein. An output derived from for example, a scanner sensor 13 is digitized to produce bit mapped image data representing the document image for each page of the document, which data is stored, for example, in a memory 15 of a special or general purpose digital computer 16. The digital computer 16 can be of the type that performs data driven processing in a data processing system which comprises execution processing means for performing functions by executing program instructions in a predetermined manner, such computers now being well known in the art. The output from the computer 16 is delivered to an output device, such as, for example, a memory or other form of storage unit, or an output display 17 as illustrated, which may be, for instance, a photocopier, CRT display, printer, facsimile machine, or the like.

In accordance with the invention, the frequency of selected image units is first determined as set forth in the aforesaid copending application of Cass et al. which is incorporated herein by reference. Thus, with reference now to FIG. 1, the first phase of the image processing technique of the invention involves a low level document image analysis in which the document image for each page is segmented into undecoded information containing image units (step 20) using conventional image analysis techniques; or, in the case of text documents, preferably using the bounding box method described in copending U.S. patent application Ser. No. 07/794,392 now U.S. Pat. No. 5,321,770, filed concurrently herewith by Huttenlocher and Hopcroft, and entitled "Method and Apparatus for Determining Boundaries of Words in Text". The locations of and spatial relationships between the image units on a page are then determined (step 25). For example, an English language document image can be segmented into word image units based on the relative difference in spacing between characters within a word and the spacing between words. Sentence and paragraph boundaries can be similarly ascertained. Additional region segmentation image analysis can be performed to generate a physical document structure description that divides page images into labelled regions corresponding to auxiliary document elements like figures, tables, footnotes and the like. Figure regions can be distinguished from text regions based on the relative lack of image units arranged in a line within the region, for example. Using this segmentation, knowledge of how the documents being processed are arranged (e.g., left-to-right, top-to-bottom), and, optionally, other inputted information such as document style, a "reading order" sequence for word images can also be generated. The term "image unit" is thus used herein to denote an identifiable segment of an image such as a number, character, glyph, symbol, word, phrase or other unit that can be reliably extracted. Advantageously, for purposes of document review and evaluation, the document image is segmented into sets of signs, symbols or other elements, such as words, which together form a single unit of understanding. Such single units of understanding are characterized in an image as being separated by a spacing greater than that which separates the elements forming a unit. Such image units representing single units of understanding will be referred to hereinafter as "word units."

Advantageously, a discrimination step 30 is next performed to identify the image units which have insufficient information content to be useful in evaluating the subject matter content of the document being processed. Such image units include stop or function words, i.e., prepositions, articles and other words that play a largely grammatical role, as opposed to nouns and verbs that convey topic information. One preferred method is to use the morphological function word detection techniques disclosed in the copending U.S. patent application Ser. No. 07/794,190 filed concurrently herewith by Bloomberg et al., and entitled "Detecting Function Words Without Converting a Document to Character Codes".

This application discusses binary images. In this context, the term "image" refers to a representation of a two-dimensional data structure composed of pixels. A binary image is an image where a given pixel is either ON or OFF. Binary images are manipulated according to a number of operations wherein one or more source images are mapped onto a destination image. The results of such operations are generally referred to as images. The image that is the starting point of processing will sometimes be referred to as the original image or source image.

A morphological operation refers to an operation on a pixelmap image (a source image), that uses a local rule at each pixel to create another pixelmap image, the destination image. This rule depends both on the type of the desired operation to perform as well as on the chosen structuring element.

Pixels are defined to be ON if they are black and OFF if they are white. It should be noted that the designation of black as ON and white as OFF reflects the fact that most documents of interest have a black foreground and a white background. The techniques of the present invention could be applied to negative images as well. The discussion will be in terms of black on white, but the references to ON or OFF apply equally well to images which have been inverted and, therefore, the roles of these two states are reversed. In some cases the discussion makes reference to a "don't care" pixel which may be either an ON or an OFF pixel.

A structuring element (SE) refers to an image object of typically (but not necessarily) small size and simple shape that probes the source image and extracts various types of information from it via the chosen morphological operation. In the attached figures that show SEs, a solid circle is a hit, and an open circle is a miss. The center position is denoted by a cross. Squares that have neither solid nor open circles are "don't cares"; their value in the image (ON or OFF) is not probed. A binary SE is used to probe binary images in a binary morphological operation that operates on binary input images and creates an output binary image. The SE is defined by a center location and a number of pixel locations, each normally having a defined value (ON or OFF). The pixels defining the SE do not have to be adjacent each other. The center location need not be at the geometrical center of the pattern; indeed it need not even be inside the pattern. A solid SE refers to an SE having a periphery within which all pixels are ON. For example, a solid 2.times.2 SE is a 2.times.2 square of ON pixels. A solid SE need not be rectangular. A horizontal SE is generally one row of ON pixels and a vertical SE is generally one column of ON pixels of selected size. A hit-miss SE refers to an SE that specifies at least one ON pixel and at least one OFF pixel.

AND, OR and XOR are logical operations carried out between two images on a pixel-by-pixel basis.

NOT is a logical operation carried out on a single image on a pixel-by-pixel basis.

EXPANSION is scale operation characterized by a scale factor N, wherein each pixel in a source image becomes an N.times.N square of pixels, all having the same value as the original pixel.

REDUCTION is a scale operation characterized by a scale factor N in a threshold level M. REDUCTION with scale=N entails dividing the source image into N.times.N squares of pixels, mapping each such square in the source image to a single pixel on the destination image. The value for the pixel in the destination image is determined by the threshold level M, which is a number between 1 and N.sup.2. If the number of ON pixels in the pixel square is greater or equal to M, the destination pixel is ON, otherwise it is OFF.

EROSION is a morphological operation wherein a given pixel in the destination image is turned ON if and only if the result of superimposing the SE center on the corresponding pixel location in the source image results in a match between all 0N and OFF pixels in the SE and the underlying pixels in the source image. An EROSION will give one pixel in the destination image for every match. That is, at each pixel, it outputs 1 if the SE (shifted and centered at that pixel) is totally contained inside the original image foreground, and outputs 0 otherwise. Note that EROSION usually refers to operations using a SE with only hits and more generally matching operations with both hits and misses (often called a hit-miss transform). The term EROSION is used herein to include matching operations with both hits and misses, thus the hit-miss transform is the particular type of EROSION used herein.

DILATION is a morphological operation wherein a given pixel in the source image being ON causes the SE to be written into the destination image with the SE center at the corresponding location in the destination image. The SEs used for DILATION typically have no OFF pixels. The DILATION draws the SE as a set of pixels in the destination image for each pixel in the source image. Thus, the output image is the union of all shifted versions of the SE translated at all 1-pixels of the original image.

FillClip is a morphological operation where one image is used as a seed and is grown morphologically, clipping it at each growth step to the second image. For example, a fillClip could include a DILATION followed by logically ANDing the DILATION result with another image.

OPENING is a morphological operation that uses an image and a structuring element and consists of an EROSION followed by a DILATION. The result is to replicate the structuring element in the destination image for each match in the source image.

CLOSING is a morphological operation using an image and a structuring element. It includes a DILATION followed by an EROSION of the image by a structuring element. A CLOSE of an image is equivalent to the bit inverse of an OPEN on the (bit inverse) background.

UNION is a bitwise OR between two images. An intersection is a bitwise AND between two images.

Blurring is a DILATION of an image by a structuring element(s) consisting of two or more hits.

A mask refers to an image, normally derived from an original or source image, that contains substantially solid regions of ON pixels corresponding to regions of interest in the original image. The mask may also contain regions of ON pixels that do not correspond to regions of interest.

The various operations defined above are sometimes referred to in noun, adjective, and verb forms. For example, references to DILATION (noun form) may be in terms of DILATING the image or the image being DILATED (verb forms) or the image being subjected to a DILATION operation (adjective form). No difference in meaning is intended.

Morphological operations have several specific properties that simplify their use in the design of appropriate procedures. First, they are translationally invariant. A sideway shift of the image before transforming does not change the result, except to shift the result as well. Operations that are translationally invariant can be implemented with a high degree of parallelism, in that each point in the image is treated using the same rule. In addition, morphological operations satisfy two properties that make it easy to visualize their geometrical behavior. First, EROSION, DILATION, OPEN and CLOSE are increasing, which means that if image 1 is contained in image 2, then any of these morphological operations on image 1 will also be contained in the morphological operation on image 2. Second, a CLOSE is extensive and OPEN is antiextensive. This means that the original image is contained in the image transformed by CLOSE and the image transformed by OPEN is contained in the original image. The DILATION and EROSION operations are also extensive and antiextensive, respectively, if the center of the structuring element is located on a hit.

The OPEN and CLOSE operations also satisfy two more morphological properties:

(1) The result of the operation is independent of the position of the center of the structuring element.

(2) The operation is idempotent, which means that reapplying the OPEN or CLOSE to the resulting image will not change it.

An image unit means an identifiable segment of an image such as a word, number, character, glyph or other units that can be extracted reliably and have an underlying linguistic structure.

The term significant and its derivatives are used in this description to indicate the importance of particular characteristics of an image unit. An image unit with significant characteristics becomes a significant image unit in that it contains high value information which can be used for further processing of the document image. Significant characteristics of image units include a variety of classifiers such as length, width, location on a page of the document image, font, typeface and measurement by other parameters including, but not limited to:

one or more cross-sections of a box (a cross-section being a sequence of ON or OFF pixels);

a number of ascenders associated with an image unit;

a number of descenders associated with an image unit;

average pixel density in an image unit;

a length of a topline contour of an image unit, including peaks and troughs;

a length of a base contouring of the image units, including peaks and troughs; and

the location of image units with respect to neighboring image units, e.g., vertical position and horizontal inter-image unit spacing.

FIG. 27 is a flow chart of a preferred embodiment of a method according to the invention for detecting function words in a scanned document image without first converting the document image to character codes. An image of a page of a document is scanned in step 602 and the image is segmented into image units in step 604. The segmentation can be accomplished, for instance, by using first a technique to determine baselines of image units and then second a technique for providing bounding boxes around image units as described in the above incorporated concurrently filed U.S. Patent Application entitled "A Method of Deriving Wordshapes for Subsequent Comparison" by Huttenlocher et al., U.S. patent application Ser. No. 07/794,391. In step 606, a length and height of each image unit in the image is determined, and short image units are determined in step 608 as image units of no more than a predetermined number of characters, preferably three characters or less in length. In step 610, image units which are not short image units are deleted from the image. In step 612, the image is blurred or smeared in a horizontal direction although the image units are not smeared together. This can be accomplished for example by CLOSING the image with a horizontal structuring element such as the structuring element of length 5 (i.e., 5 pixels) as shown in FIG. 5. The length of the horizontal structuring element used to blur the x-height characters in the image is dependent upon the width of the character type being used. Furthermore, other configurations of structuring elements may be used in the CLOSING operation to obtain the same smearing effect. However, the most efficient and effective way to smear characters of x-height is to use a horizontal structuring element as described above.

A UNION of erosions is taken in step 614 of the image using a set of ascender matching structuring elements such as those shown in FIGS. 29A-1, 29A-2, 29B-1, 29B-2, and 29C-1, 29C-2, and, a set of descender matching structuring elements such as those shown in FIGS. 30A-1, 30A-2, 30B-1, 30B-2, 30C-1 and 30C-2. The UNION taken in step 614 provides optional noise elimination filtering, and the UNION will provide a seed from which to fill short image unit masks in a subsequent seed filling operation such as the fillClip operation of step 616. The UNION of step 614 acts on all image units remaining in the image (i.e., only short image units in this case) and since the UNION of erosions was taken using a set of ascender matching structuring elements and a set of descender matching structuring elements, the image units that will be filled are those containing ascender and/or descender characters, i.e., function words. The function words are identified in step 18 as those image units which are filled short image unit masks. In step 620 a test occurs to determine whether a last page of the document has been scanned. If the last page has been scanned, then the method terminates at step 624, otherwise the page is incremented in step 622 and the incremented (next) page is scanned in step 602 whereupon the image (next page) is scanned and the previously described steps of the method are reiterated. Of course, all pages could be scanned and stored as bit map images in a memory prior to performing the function word identification procedures described above. Moreover, the image segmentation step can also be performed prior to practicing the present invention (and the segmented image stored in memory) since segmentation has utility for procedures other than with the present invention.

FIG. 28 is a block diagram of a preferred embodiment according to the invention of an apparatus 650 for detecting function words in a scanned document image without first converting the document image to character codes. The apparatus 650 includes a user input device 652 which includes, for example, one or more of an image scanner (optical or electronic) a keyboard, a touchscreen, a mouse, etc. The image scanner can be a stand-alone device, or part of a facsimile machine or document reproducing machine (copier). A processor 656 is connected to the input device 652 for processing the document image to segment the document image into image units, determine a length and height of each image unit, determine short image units as image units of preferably three characters or less in length, delete image units from the image that are not short image units, blur the image in a horizontal direction, take a UNION of erosions of the image using a set of ascender and descender matching structuring elements, fillClipping the image to provide short image unit masks filled for short image units with at least one ascender and short image unit masks filled for short image units with at least one descender, and identifying function words as those image units which are filled short image unit masks. Processor 656 operates to perform these functions in accordance with operating programs read from read only memory (ROM) 660, and by using random access memory (RAM) 662. Documents can be stored or retrieved from files 664, and processing results and user input information can be monitored on a CRT display monitor 654.

Next, in step 40, selected image units, e.g., the image units not discriminated in step 30, are evaluated, without decoding the image units being classified or reference to decoded image data, based on an evaluation of predetermined image characteristics of the image units. The evaluation entails a determination (step 41) of the image characteristics and a comparison (step 42) of the determined image characteristics for each image unit with the determined image characteristics of the other image units.

One preferred method for defining the image unit image characteristics to be evaluated is to use the word shape derivation techniques disclosed in copending U.S. patent application Ser. No. 07/794,391 filed concurrently herewith by Huttenlocher and Hopcroft, and entitled "A Method for Deriving Wordshapes for Subsequent Comparison," Published European Application No. 0543594, published May 26, 1993. As described in that application, at least one, one-dimensional signal characterizing the shape of the word unit is derived; or an image function is derived defining a boundary enclosing the word unit, and the image function is augmented so that an edge function representing edges of the character string detected within the boundary is defined over its entire domain by a single independent variable within the closed boundary, without individually detecting and/or identifying the character or characters making up the word unit. As part of this process, baselines of the text on a page are determined (a baseline is a line extending under the non-descending characters on a text line). It will be appreciated that the ordering of the word units along the baselines, and the ordering of the baselines on each document image page provides the reading order of word units in the document image.

More specifically, this application discloses as illustrated in FIG. 3A, a generalized image processing system, which covers numerous situations in which the present invention may find advantageous use. Generally, a source image may be derived from a source image derivation system 2, which may be a scanner, facsimile device, or storage system. The source image is forwarded to a computer processing device 4 which may be any of several well known devices including the inventive device described herein. In response to commands entered at user interface 6, processing device 4 produces an output at an output device 8a, which may be a printer, display, facsimile device or other storage device. In essence, as is shown in the upper portion of FIG. 3A and 3B, an input document is directed into a system and an output document is retrieved from it.

In the following description, an image is generally described as an image bitmap, where an image is represented as a plurality of image signals. These signals, commonly referred to as pixels, are typically denoted as black when intended to represent a corresponding mark or active position on a document from which they were produced. However, these constructs have been used to enable the description of the present invention, and are in no way intended to limit the domain of such to that of black-and-white or binary images. Rather, the present invention is generally applicable across a broad range of image representation techniques.

FIG. 3B, shows a system which embodies the present invention for deriving, defining, and comparing words in terms of their shapes. It will, of course, be recognized that each element of the system may be many devices, or may simply be a program operated within a single device. Beginning with an input bitmap 10a, whose source is indeterminate, and not part of the invention, a bitmap of an image is initially directed to a segmentation system 12a, in which words, or character strings, or other multi-character units of understanding, will be derived. Initially, the image bitmap passes through skew detector 14, which determines the angle of orientation of text in the image. Using information about the orientation of the image, and the image itself, at text baseline processor 16a, toplines and baselines of the text are determined, so that upper and lower boundaries of lines of text within the image are identified. At median filter 18, the function referred to as "blobify" is performed, which operates on the image so that each word group in a line may be treated as a single unit. As used herein, "word" " symbol string" or "character string" refers to a set of connected alphanumeric or punctuation elements, or more broadly, signs or symbols which together form a single unit of semantic understanding. It will be appreciated that these terms may also be used to refer to the images thereof. Such single units of understanding are characterized in an image as separated by a spacing greater than that which separates the elements, signs or symbols forming the unit. To the blobified image, a set of white lines are added at block 20a, to clearly separate adjacent lines of text. The white lines are based on baseline determinations provided by processor 16a. Using this information, i.e., the blobified words, which are clearly separated from adjacent words and words in adjacent lines, a bounding box is defined about the word at block 22, thereby identifying and enclosing the word.

Thereafter word shape signal computer 24 derives a word shape signal representing the individual words in the image, based on the original image and the bounding box determinations. This information is then available for use at a word shape comparator 26, for comparing word shape signals representative of known words from a word shape dictionary 28, with the as yet unidentified word shape signals. In an alternative embodiment word shape comparator 26 may be used to compare two or more word shapes determined from image 10a. More importantly, word shape comparator 26 is not limited to the comparison of word shapes from unrecognized strings of characters to known word shapes. In a simplified context, comparator 26 is merely an apparatus for comparing one word shape against another to produce a relative indication of the degree of similarity between the two shapes.

In general, a method accomplishing the invention includes the following steps. Once orientation of the image is established and line spacing and word group spacing is established, each word can be surrounded by a bounding box. A reference line is then created extending through the character string image. The reference line may be a block having a finite thickness ranging from two-thirds of the x height to one-third of the x height, or in fact it may have a zero width. At the resolution of the image, the distance from the reference line to the upper edge of the text contour or bounding box is measured in a direction perpendicular to the reference line. Similarly, measurements may be made from the reference line to the lower bounding box edge or to the text contour along the lower portion of the word, whichever is closer. Because the set of values derived computationally can be expressed in terms of position along the horizontal axis versus length, the signal can be considered a single independent variable or one dimensional signal. Either or both of these sets of values may be used to describe the word shape. Additionally, although possibly less desirable, it is well within the scope of the invention to measure the distance of a perpendicular line drawn from the top of the bounding box or the bottom of the bounding box, to the first contact with the word or the reference line, as desired.

With a system and process for word shape derivation given, the invention may also be considered mathematically. Considering image data i(x,y), which in one common case could be an array of image data in the form of a bitmap, a character set is identified in one of many methods, perhaps as described above, which defines a boundary enclosing the selected symbol string within a subset of the array of image data. From i(x,y), an edge signal e(x,y), which represents the edges of i(x,y) detected within the closed boundary, is derived. The edge signal is augmented by adding additional data to i(x,y) so that e(x,y) is a signal e'(x,y) defined over its entire domain with respect to a single dimension or variable within the closed boundary. One, two, or more signals may be derived from e'(x,y) which are each one dimensional signals g'(t), where g is a function of parameter t which is a reference frame dependent parameter.

It is important to realize that the mathematical process used for the derivation of the one dimensional signal is essentially reversible up to the information it contains, e.g., a bitmap may be reconstructed from the upper and lower bitmap contours, as illustrated in FIG. 18. It will be noted that if the reference has a finite thickness and is therefore taken out of the image, that portion of the image is not identifiable, however, if it has a zero width the information still remains.

A recognition dictionary, or look up table of word shapes, can clearly be created through use of the described process. The process can be operated on using either scanned words as the source of the information, or in fact, they can be computer generated for a more "perfect" dictionary.

To demonstrate the process of the invention, at FIG. 4, a sample image, taken from a public domain source is shown, having several lines of text contained therein. FIG. 4 demonstrates approximately how the image would appear on the page of text, while FIG. 5, shows a scanned image of the page, which demonstrates an enlargement of the image of a bitmap that would present problems to known OCR methods. Looking at, for example, the image of the word 50a "practitioner" in the first line of the text image, it may be seen that several of the letters run together. Also, at the lower right hand portion of the image, circled and numbered 52, noise is present. Looking at the word "practitioner's", circled and numbered 54, the running together of a punctuation mark and a letter is further noted.

With reference again to FIG. 3B, in one possible embodiment of the invention, skew detector 14, may be implemented using a general method for determining the orientation of the text lines in the image. This method looks at a small number of randomly selected edge pixels (defined as a black pixel adjacent to at least one white pixel), and for each edge pixel considers, at FIG. 6A, a number of lines, 56a, 56b, 56c being examples, extending from the pixel at evenly spaced angular increments over a specified range of angles. The edge pixels are selected randomly from the set of all image pixels by the function RandomEdgePixel() (Appendix, page 243). FIGS. 6A (see lines 56a, 56b, 56c), 6B (see lines 58a, 58b, 58c) and 6C (see lines 60a, 60b, 60c) represent a series of increasingly smaller angular ranges over which the above mentioned technique is applied to illustrative edge pixels to accurately determine the angular orientation of the text within the image. Subsequent to finding edge pixels and defining the lines, skew detector 14 traces the path of each line, determining the lengths, in pixels, of strings of successive black pixels which are intersected by the line. Upon reaching the image boundary, an average black pixel string length is calculated by summing the lengths of the individual strings, and dividing the sum by the total number of distinct strings which were found. This operation is carried out for all the lines, thereby arriving at an average black pixel string length for each line extending from the selected edge pixel. These lengths are plotted on FIG. 6D as curve A, showing minima at approximately 0 and 3.14 radians. Curve A is a graphical representation of the summation/averaging function over each of a series of angled lines extending from the edge pixel, and spread over a range from 0 to 2.pi. radians. Once a first minimum has been located, verification of the minimum (in the example, approximately 0 radians) is achieved by determining whether a second minimum exists at approximately .pi. radians from the first minimum. Upon verifying the existence of a second minima (in the example, approximately 3.14 or .pi. radians), a coarse skew angle is identified. Subsequently, it is necessary to more closely determine the skew angle of the text. This is accomplished by utilizing a number of lines which extend from a randomly selected edge pixel, where the lines differ by smaller angular increments, and the angular range is centered about the coarse skew angle. However, the fine skew angle may be determined by analyzing the total number of black pixels contained along a predetermined length of the lines. More specifically, the number of pixels over a unit distance are plotted as curve B on FIG. 6D, and the fine skew angle is determined by identifying the maxima of the curve. In other words, the point of the curve where the highest concentration of black pixels per unit line length exists, more accurately represents the angle of the text lines in the image. As shown by curve B, this results in a fine skew angle of approximately 0 radians, where the line intersects with the most black pixels along its length, and therefore is representative of the closest angle of orientation that needs to be determined.

Alternatively, the skew angle may be determined as indicated by the NewFine() function (Appendix, page 245), which determines the skew angle using multiple iterations of the procedure described with respect to the fine angle determination. As indicated by FIGS. 6A, 6B, and 6C, each iteration would also use lines covering an increasingly smaller angular range, until a desired skew angle accuracy is reached. In the implementation illustrated by FIGS. 6A, 6B, and 6C, the desired accuracy is achieved by a series of three iterations, each using a series of 180 distinct angles about the selected edge pixel.

In the next process step, illustrated in the graphs of FIG. 7A and FIG. 7B, text baseline processor 16a identifies the characteristic lines, upper topline and lower baseline, of each line of text. The process steps executed by text baseline processor 16a are illustrated in detail in FIG. 8A and 8B. The histogram of FIG. 7A, shown to the left along the image, is derived by examining lines, at the resolution of the image, and oriented parallel to the skew orientation of the image, as defined by the previously determined skew angle. These parallel lines spanning the image are used to determine the number of black pixels intersected by each of the lines. Along lines passing through inter text line spaces, no black pixels should be intercepted, while along lines through the text, large numbers of black pixels should be intercepted.

More specifically, the function BaseLines(), (Appendix page 160), first finds the coordinates of a "main" line, block 142, constructed through the center of the image and perpendicular to the text lines, as determined by the skew angle passed to the function as shown by block 140. Next, Line Engine Procedure 144 is executed, where by proceeding along the main line from one end to the other, at a series of points along the main line, perpendicular branch lines are constructed which extend outwardly from the main line for a fixed distance, block 146. Along the branch lines, the number of black vertical edge pixels are counted, block 148, and the number of black pixels intersected by the lines are counted, block 150, and summed for the opposing pairs of lines, block 152. Black vertical edge pixels, as counted by block 148, are defined as black pixels having a white neighboring pixel at either the upper or lower neighboring pixel position. LineEngine()procedure 144 is repeated until all points, and associated branch lines, along the main line have been processed, as determined by decision block 154. An x-height value may be returned from this procedure, which will subsequently be used by the word shape computer 24.

Subsequently, the counts for all the branch lines are analyzed to determine the branch line pairs having the highest ratio of black vertical edge pixels to black pixels. In general, those lines having the highest percentages would correspond to lines passing along the upper and lower edges of the characters which form the text lines. As illustrated in the enlarged view of FIG. 7B, a definite distinction exists between those branch lines having a high vertical edge pixel ratio, line 82, and those having a low ratio, line 84. Application of a filter mask and comparison of the maximum peaks within the mask enables the identification of those lines which represent the text toplines and baselines, for example, line 82. The process is implemented in the maxFilter.c module, beginning at line 57, the code for which is also incorporated in the newBaselines.c module at line 274, page 214. Baseline determination is described in further detail in a copending U.S. Patent Application, for a "Method for Determining Boundaries of Words in Text", Huttenlocher et al., U.S. patent application Ser. No. 07/794,392, now U.S. Pat. No. 5,321,770 which has been previously incorporated herein by reference. An additional test may also be applied to the histogram operation of step 150. This added test, a boolean test, may be used to assure that a minimum run of black pixels was detected during the analysis of the line. For example, a flag, which is cleared at the start of each branch line analysis, may be set whenever a series of five sequential black pixels are detected along the line. This test would assure that small noise or image artifacts are not recognized as baselines due to a high vertical edge pixel ratio.

As an alternative method, it is possible to utilize the total number of black pixels lying along the branch lines to determine the locations of the baselines. Using histogram curve BL, which represents the number of black pixels counted along the branch lines, it is possible to determine which branch lines have the most black pixel intersections. Applying a threshold of the maximum allows the determination of the upper and lower characteristic line pairs for each text line. Hence, the rising and falling portions of the histogram curve BL, constitute the characteristic lines of the text, and the threshold would be used to specifically identify the localized maxima surrounding an intervening minima, thereby enabling identification of the baseline positions which would be used for further processing. More importantly, this alternative approach, illustrated as step 162, may be utilized to identify the upper and lower baselines of a baseline pair, based upon the slope of the BL histogram curve. It is important to note that there is little additional processing associated with the identification step as the histogram information was collected previously during step 150. Once the preliminary characteristic line or baseline pairs are identified, block 162, a verification step, block 164, is executed to verify that the baseline pairs are separated by more than a minimum distance, the minimum distance being established by calculating the average line pair separation for all line pairs in the image. After verification, the valid baseline information is stored by output block 166 for later use by the white line addition and segmentation blocks, 18 and 20a, respectively.

An important advantage of these baseline determination methods, are that they are highly insensitive to noise or extraneous marks in the interline space. FIG. 9 shows the result of the baseline determination on the example image of the sample text, showing that baseline pair, baseline and topline B.sub.n and B.sub.n ', respectively, have been located on the image, indicating those portions of the image in which a predominant portion of the text occurs. While some portions of the character ascender strokes are outside the baselines, no detriment to the remainder of the process is noted. Of course, a smaller threshold value might enable the system to capture more of the ascending strokes.

With reference again to FIG. 3B in conjunction with FIGS. 10 and 11, the next process step is a word group isolation step. A filter 18 is applied to a copy of the image which results in an image that tends to render the word into blobs distinguishable from one another. The filter is applied with a small window, to each area, to render as black those areas that are partly black. As shown in FIG. 10, the blobify function (Appendix page 165) first initializes mask variables which establish the mask size and angle, block 180, and then processes the upper scanline to initialize the data array, block 182. Median filtering is accomplished by sequentially moving the mask window through the image, blocks 184 and 186, and whenever the number of black pixels appearing in the window exceeds a threshold value, the target pixel, about which the window is located, is set to black. FIG. 11, which illustrates some examples of the filter process, has a mask window 200 placed over a portion of the image. For example, with a twenty percent threshold and a generally rectangular mask having twenty-one pixels, arranged at an angel approximately equal to the skew determined for the text, the result of filtering in window 200 would be the setting of pixel 204 to black. Similarly, window 206, which primarily lies within the intercharacter spacing between the pixel representations of the letters "r" and "o", would cause pixel 208 to be set to black. On the other hand, window 210, which lies in the region between word groups, would not have a sufficient number of black pixels present within the window to cause pixel 212 to be set to black. The size, shape and orientation of mask window 200 is optimized to reduce the filling in between text lines, while maximizing the fill between letters common to a single word.

As illustrated by FIG. 12, the result of the median filtering is that the relatively small spacing between characters in a word generally becomes inconsequential, and is filled with black pixels. Words become a single connected set of pixels, i.e., no white spaces completely separate characters in a single word. However, the relatively large spacing between character strings or between words, is a larger space outside of the ability of the filter to turn into black, and therefore serves to distinguish adjacent symbol strings. With reference now to FIGS. 9 and 12, it can be seen that the first two words of the sample text, "A" and "practitioner" have been "blobified", as this process is referred to, so that, for example, the "p" of "practitioner" is no longer separated from the "r" of that word. (Compare, FIG. 5). Once again, despite the blobifying or blurring of characters, "A" and "practitioner" remain as discrete blobs of connected symbols, or words.

With reference again to FIG. 3B, as an adjunct to this step, white line addition 20a, superimposes upon the blobified image of FIG. 12 a series of white pixel lines to make certain that lines of text are maintained separately from adjacent lines of text (i.e., no overlapping of the filtered text lines). With reference to FIGS. 12 and 13, noting the circled areas 258 and 258', a combination of an ascender and descender has resulted in an interline merging of two words. The text line overlap illustrated in area 258 of FIG. 12 is exactly what is eliminated by superimposing the white lines on the blobified or filtered image.

This superposition of white lines operation, the outcome of which is illustrated by FIG. 13, is carried out by the process illustrated in FIG. 14 as executed in the DrawMiddleLines() function (Appendix page 233). Generally, white lines WL are added to the image, approximately halfway between adjacent baseline and topline pairs, to assure that there is no cross-text line blobifying. Once again, FIG. 13 shows the result of white line addition to the blobified image of FIG. 12.

Referring now to FIG. 14, white line addition block 20a begins by initializing variables in step 280 and subsequently reads in the topline location from the baseline information of the first text line. The topline information is discarded, block 282, and the next baseline and topline locations are popped from the storage stack or list, blocks 284 and 286, respectively. With respect to the image, this baseline-topline pair respectively represents the bottom and top of adjacent text lines. Next, at step 288, the point lying at the center of the pair is located to provide a starting point for the white lines which are drawn from the center of the image in an outward direction. The endpoints of the white lines are calculated in step 290, using the skew angle determined by skew detector I 4 of FIG. 3B. White lines are drawn or superimposed on the blobified image at step 292, and the process is continued until all text lines have been effectively separated, as controlled by test block 294.

With reference again to FIG. 3B, as a result of the blobify or median filtering, the position of bounding boxes about each connected set of pixels formed in the blobify step may be determined. Bounding boxes are placed only about those connected components or words that are in a text line lying between the superimposed white lines. The bounding boxes are placed at the orientation of the text line, by identifying the extreme points of each group of connected pixels in the direction of the text line, and in the direction orthogonal to the text line, as opposed to the image coordinate system. This operation is performed by the function FindBorders(), (Appendix, page 172). Generally, the FindBorders function steps through all pixels within the image to find the bounding boxes of the connected characters (Paint Component), to determine the coordinates of the upper left corner of each box, as well as the length and width of the box.

Referring now to FIGS. 15A and 15B, which detail the FindBorders() procedure, segmentation step 22 begins by placing a white border completely around the filtered image, step 300. This is done to avoid running outside the edge of the array of image pixels. Next, pixel and line counters, x and y, respectively, are initialized to the first pixel location inside the border. Calling the ReadPixel procedure, block 304, the pixel color (black or white) is returned and tested in block 306. If the pixel is white, no further processing is necessary and processing would continue at block 322. Otherwise, the PaintComponent() procedure (Appendix, page 171) is called and begins by storing the location of the black pixel in a queue, block 308. Subsequently, in a copy of the image, the pixel is set to white and the boundaries of the box, surrounding the connected pixels or components, are updated, blocks 310 and 312, respectively. Next, adjoining black pixels are set to white, block 314, and the locations of the black pixels are added to the end of the queue, block 316. At block 318 the queue pointers are tested to determine if the queue is empty. If not empty, the next pixel in the queue is retrieved, block 320, and processing continues at block 312. Otherwise, if the queue is empty, all of the connected black pixels will have been set to white and the box boundaries will reflect a box which encompasses the connected components. Subsequently, the boundaries of the box which encompasses the word segment are verified and may be adjusted to an orthogonal coordinate system oriented with respect to the skew of the text lines, block 322.

It will no doubt be apparent here that while finding each text line is an integral part of the described method, and serves to make the present embodiment more robust, other methods of deriving the information acquired by that step are possible. The primary use of the text line finding function is a) to determine x-height, and b) define the white line addition for separating interline blobs. Certainly this step may be removed, with a sacrifice in robustness, or other means of deriving the necessary information may be available.

The looping process continues at block 324 which checks pixel counter x to determine if the end of the scanline has been reached, and if not, increments the counter at block 326 before continuing the process at block 304. If the end of the scanline has been reached, pixel counter x is reset and scanline counter y is incremented at block 328. Subsequently, block 330 checks the value of scanline counter y to determine if the entire image has been processed. If so, processing is completed. Otherwise, processing continues at block 304 for the first pixel in the new scanline.

Thus, as shown in FIG. 16, for the word "practitioner" the extremities of the connected character image define the bounding box. Once bounding boxes have been established, it is then possible at this step, to eliminate noise marks from further consideration. Noise marks are determined: 1) if a bounding box corner is outside the array of image pixels (Appendix, page 171); 2) if a box spans multiple text lines in the array (Appendix 229), or lies completely outside a text line; 3) if boxes are too small compared to a reference .epsilon., in either or both longitudinal or latitudinal directions, and accordingly are discarded. Noise marks 70a and 72 and others will not be considered words. The OnABaseline() function (Appendix, page 229) is an example of a function used to eliminate those boxes lying outside of the baseline boundaries.

With reference to FIG. 3B, at word shape computer 24, a signal representing the image of a word, or at least a portion thereof, now isolated from its neighbors, is derived. The derived signal is referred to as a word shape contour. The shape contour for each word is determined using the MakeShell() function (Appendix, page 228). As illustrated in FIG. 17A, this function first moves along the top of each bounding box, and starting with each pixel location along the top of the box, scans downward relative to the page orientation, until either a black pixel, or the bottom of the box, is reached. A record of the set of distances d between the top of the box and the black pixel or box bottom is maintained. The set of distances d, accumulated over the length of the box, constitutes the top raw contour of the word shape. Subsequently, a bottom raw contour is produced in a similar manner as illustrated in FIG. 17B, for the same word depicted in FIG. 17A, by sequentially moving across the bottom of the box, and looking in an upwards direction, for either the first black pixel or the top of the bounding box.

With reference now to FIG. 19, at block 100 which preferably operates on the actual image as opposed to the filtered image, which could be used in this step, one or more reference lines are established through each word. In other terms, the data representing the symbol string is augmented, so that it is defined over the range of the symbol string. In one embodiment, a blackout bar, which may have a finite thickness or a zero thickness is constructed through the word, preferably having an upper limit or reference line at approximately two thirds of the x height, and a lower limit or reference line at approximately one-third of the x height (which was determined at the baseline determination step). At contour calculation 102, a set of measurements is derived, for the distance d between the upper or lower edge of the bounding box, and the word, or the nearer of the reference line's closer edge of the black out bar. The calculation's measurements are made at the resolution of the image. With reference to FIG. 20A, where the calculation's measurements are illustrated pictorially, it can be seen that the reference lines serve to allow the signal that will ultimately be derived from this step to be defined at every sampling position over the length of the word. In a preferred embodiment, the calculation's measurements of d are actually generated from the contour data derived in accordance with FIGS. 17A, 17B previously collected, and are adjusted to limit the distance d with either the upper or lower edge of the blackout bar as indicated. In the embodiment shown in FIG. 20A, measurements are made from the upper line of the bounding box to the upper reference line of the word, although this is not a requirement. Thus, for example, the measurement could alternatively be made from the reference line to either the upper or lower bounding line, or the character. FIG. 20B better shows how the set of measurements is used to form the signal output from block 104. The contour is represented as a set of measurements distance d', relative to the reference line. Measurement d' is therefore derived from the measurements shown in FIG. 20A, which designate the stopping point of line d, and the known position of the black out bar. Calculating the distance relative to the reference line enables scaling of the word shape contours to a common x height, thereby facilitating any subsequent comparison of the shapes. Accordingly, the distances d' represent a measurement from the reference line or blackout bar to the outer limits of the letter, and in the absence of a letter, provide a zero measurement. These measurement might be derived directly, but the proposed indirect methods appear easier to implement. FIGS. 20C and 20D show that the sets of d' values can be plotted on a graph to form a one dimensional signal or curve representing the word shape. Details of the contour determination are contained in the function StoreOutlinePair() beginning in the Appendix at page 255. FIG. 18 is an image of the contour locations as established for the text sample of FIG. 4. It is important to note the informational content of FIG. 18, where, for the most part, it is relatively easy to recognize the words within the passage by their contours alone.

In studies of the information delivered by the appearance of English language words, it has been determined that in a majority of cases, words can be identified by viewing only approximately the top third of the image of the word. In other words, the upper portion of the word carries with it much of the information needed for identification thereof. In a significant portion of the remainder of cases, words that are unidentifiable by only the upper third of the image of the word, become identifiable when the identification effort includes the information carried by the lower third of the image of the word. A relatively small class of words requires information about the middle third of the word before identification can be made. It can thus be seen that a stepwise process might be used, which first will derive the upper word shape signal or contour, second will derive the lower word shape signal or contour, and thirdly derive a word shape signal central contour (from the reference line towards the word or bounding box), in a prioritized examination of word shape, as required. In the examples of FIG. 20A, 20B, and 20C, the word "from" is fairly uniquely identifiable from its upper portion only. In the examples of FIG. 21A, 21B, 21C and 21D, the word "red" is less uniquely identifiable from its upper portion, since it may be easily confused with the word "rod" and perhaps the word "rad". While the lower portion of the letter "a" may distinguish "red" and "rad", it is doubtful that the lower portion of the letter "o" will distinguish the words "red" from "rod". However, the central portions of "red", "rad", and "rod" are quite distinct.

The determined morphological characteristic(s) or derived image unit shape representations of each selected image unit are compared, as noted above (step 41), with the determined morphological characteristic(s) or derived image unit shape representations of the other selected image units for the purpose of identifying equivalence classes of image units (step 50), such that each equivalence class contains most or all of the instances of a given word in the document. The equivalence classes are thus formed by clustering the image units in the document based on the similarity of image unit classifiers, without actually decoding the contents of the image units, such as by conversion of the word images to character codes or other higher-level interpretation. Any of a number of different methods of comparison can be used. One technique that can be used, for example, is by correlating the raster images of the extracted image units using decision networks, such technique being described for characters in a Research Report entitled "Unsupervised Construction of Decision networks for Pattern Classification" by Casey et al., IBM Research Report, 1984, said Research Report being incorporated by reference herein.

Preferred techniques that can be used to identify equivalence classes of word units are the word shape comparison techniques disclosed in U.S. patent application Ser. Nos. 07/796,119 and 07/795,169, filed concurrently herewith by Huttenlocher and Hopcroft, and by Huttenlocher, Hopcroft and Wayner, respectively, and entitled, respectively, "Optical Word Recognition By Examination of Word Shape," Published European Application No. 0543592, published May 26, 1993 and "Method for Comparing Word Shapes". For example, U.S. patent application Ser. No. 07/795,169 discloses, with reference to FIG. 3B, one manner in which a comparison is performed at word shape comparator 26. In one embodiment, the comparison is actually several small steps, each of which will be described. With reference to FIG. 22, generally, the two word shape signals, one a known word, the other for an unknown string of characters are compared to find out whether they are similar. However, in this case, signal R is the upper contour of the word "red", while signal F is the upper contour of the word "from". Actually, relatively few signals could be expected to be exactly identical, given typical distinctions between character fonts, reproduction methods, and scanned image quality. However, the word shape signals to be compared may be scaled with respect to one another, so that they have the same x-heights. This is achieved by determining the x-height of the pair of word shape contours to be compared. Once determined, the ratios of the x-heights are used to determine a scale factor to be applied to one of the contours. As the x-height is a characteristic measurement for fonts, it is used to determine the scaling factor in both the horizontal and vertical directions. An example of the scaling operation is found in the fontNorm.c file beginning at line 172, where the StoreOutlinePair() function carries out the scaling operation in both the x and y, horizontal and vertical, directions. Alternatively, the shape signals may be compared without normalization and a weighing factor imposed upon the portion of the measured difference due to the unequal lengths. Furthermore, the amplitude or height of the signals has been normalized to further reduce the impact of the font size on the word shape comparison.

Referring next to FIGS. 23A-23C, which illustrate details of the ascender/descender normalization operation, each of the shape signals are normalized based upon a common relationship between the ascender and descender heights and the x-height of the text characters. As illustrated, the actual ascender heights of characters printed with supposedly similar font size, or what is now an appropriately scaled font size, may be slightly different. This occurs as a result of type faces or fonts which are small on body or large on body, implying that similar characters exhibit variations in height across fonts that are the same size, for example 24 point fonts. As an illustration, distance d.sub.1 in FIG. 23A represents the difference in ascender height for two occurrences of the letter "h." Likewise, distance d.sub.2 illustrates a similar difference between the heights of the letter "f" in FIG. 23B. As illustrated in FIG. 23C, the typical character may be broken into three sections, ascender portion 390, x-height portion 392, and descender portion 394. In addition, the relative heights of these sections are illustrated as c, a, and b, respectively. Again, the normalization operation applied to the shape contours is found in the fontNorm.c module, beginning at page 183 of the Appendix. Applying the operations described with respect to StoreOutlinePair() function, page 255 of the Appendix, the areas of the contour lying above the x-height are scaled as follows: ##EQU1## Similarly, the descenders are scaled by the following equation: ##EQU2## where, in both cases, the value used in the numerator (1.5) is arrived at based upon observation of the relationship between ascender or descender heights and the x-height. Also included within the StoreOutlinePair() function is an operation to remove the portions of the contours which do not represent portions of the text string. These regions lie at the ends of the bounding boxes illustrated in FIG. 16. For example, the box surrounding the word "practitioner" in FIG. 16 can be seen to extend beyond the actual word image. As further illustrated at the ends of the word "from" in FIGS. 20A-20D, the contour does not contain useful information. By removing these regions from the contour shape, less error will be introduced into the comparison operations.

Subsequent to the normalization operation, standard signal processing steps can be used to determine the similarity or dissimilarity of the two signals being compared. Alternatively, the following equation may be used: ##EQU3## where .DELTA..sub.string is the difference value between the two signals;

f(x) is the known signal; and

g'(x) is the unknown signal.

In a simple determination, the difference could be examined and if it is close to zero, such would be indicated that there would be almost no difference between the two signals. However, the greater the amount of difference, the more likely that the word was not the same as the word to which it was being compared.

It is important to note that the embodiments described herein, as supported by the code listings of the Appendix, compare the word shape contours using the upper and lower contours for each word in conjunction with one another. This is an implementation specific decision, and is not intended to limit the invention to comparisons using only the top and bottom contours in conjunction with one another. In fact, sufficient information may be contained within the upper contours alone so as to significantly reduce the requirements for a comparison of the lower contours, thereby saving considerable processing effort.

The steps of this simplified comparison method, as first contemplated, are illustrated in FIG. 24. Beginning at step 410, the contour for the first word shape is retrieved from memory, and subsequently, the second word shape is retrieved by step 412. Next, the centers of gravity of the word shapes, defined by the upper and lower contours, are determined and aligned, step 414. The purpose of this step is to align the centers of the word contours to reduce the contour differences that would be attributable solely to any relative shift between the two sets of contours being compared. The center of gravity is determined by summing the areas under the curves (mass) and the distances between the contours (moments) which are then divided to give an indication of the center of gravity for the upper and lower contour pair. Once determined for both sets of contour pairs, the relative shift between the pairs is determined, step 416, and the contours are shifted prior to calculating the difference between the contours. The shifting of the contours is necessary to reduce any error associated with the establishment of the word shape boundaries and computation of the word shapes at block 24 of FIG. 3B. Step 418 handles those regions lying outside the overlapping range of the shifted contour pairs, determining the difference against a zero amplitude signal in the non-overlapping regions. This is done by summing the squared values of the upper and lower contours at the non-overlapping ends of the contours. Subsequently, the overlapping region of the contours are compared, step 420. The difference in this region is determined as the sum of the squared differences between the upper curves and the lower curves, as shown in the function L2Norm() on page 100 of the Appendix. Next, the values returned from steps 418 and 420 are added to determine a sum of the differences over the complete range defined by the shifted contours. This value may then be used as a relative indication of the similarity between the contour pairs for the two word shapes being compared.

An alternative to the center-of-gravity comparison method, uses a signal processing function known as time warping, as described in the article "Performance Tradeoffs in Dynamic Time Warping Algorithms for Isolated Word Recognition", by Myers, Rabiner, and Rosenberg, IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. ASSP-28, No. 6, December 1980, and the book, "Time Warps, String Edits, and Macromolecules: The Theory and Practice of Sequence Comparison", by Sankoff and Kruskal, Addison-Wesley Publishing Company, Inc., Reading, Mass., 1983, Chapters 1 and 4, and may be used to provide for compression and expansion of points along the contours until the best match is made. Then a score is derived based on the amount of difference between the contours being compared and the stretching required to make the contours match. Once again, the score provides a relative indication of the match between the two signals being compared.

Referring now to FIG. 25, which depicts the general steps of the dynamic warping method, the method relies on the use of a difference array or matrix to record the distances between each point of the first contour and points of the contour to which it is being compared. As illustrated in the figure, and detailed in the code listings contained in the Appendix, the process is similar for all of the measures which may be applied in the comparison.

First, the organization of the code is such that a data structure is used to dynamically control the operation of the various comparison functions. The structure DiffDescriptor, the declaration for which is found on page 9 of the Appendix (see diff.h), contains variables which define the measure to be applied to the contours, as well as, other factors that will be used to control the comparison. These factors include: normalization of the contour lengths before comparison; separate comparisons for the upper and lower contours; a centerWeight factor to direct the warping path; a bandwidth to constrain the warp path; a topToBottom ratio which enables the top contour comparison to be weighted more or less with respect to the bottom contour comparison; and a hillToValley ratio to selectively control weighing the contour differences when an unknown contour is being compared to a known or model word shape contour. Interpretation of the various factors is actually completed in the diff2.c module at page 56 of the Appendix, although descMain.c at page 49 provides an illustration of the interpretation of the factors.

In general, each measure implements a comparison technique, however, each is optimized for a specific type of dynamic comparison, for example, a slope limited dynamic warp having a non-unitary centerweight and a topToBottom weight greater than one. The first level of selection enables the use of a slope-constrained warping function for comparison, an unconstrained warp, or a simple, non-warped, comparison. Within both of the warp comparison methods, there are both separate comparison functions, where the top and bottom contours are warped independently, and parallel comparison functions, where the warp is applied to both the top and bottom contours simultaneously. Specific details of the comparison functions are generally contained within the newMatch.c file beginning at page 101 of the Appendix.

In the general embodiment, the dynamic warping process starts by allocating space for the path/distance array, step 450, which will hold the distance values generated during the comparison and warping of one word shape contour with respect to another. After allocating space, the border regions of the array must be initialized as the process used by all the warping measures is an iterative process using data previously stored in the array for the determination of the cumulative difference between the contours. At step 452, the array borders are initialized. Initialization of the first row of the array entails the determination of the square of the difference between a first point on the first contour and each point on the second contour. Subsequent to border initialization, the column and row index values, L1 and L2, respectively, are reset to 1 to begin processing the individual, non-border, points along the contours.

Processing of the contours proceeds at steps 458 through 64, where the difference in distance between each point along the second contour, with respect to a point on the first contour is calculated. Moreover, this difference, or distance, is calculated and then summed with a previously determined difference value. In addition, some of the previously determined difference values may be weighted differently, for example, in one embodiment weights of the difference values along the array diagonal may be modified by a centerWeight weighing factor. As an illustration, the operation of the NewMatch() function, beginning at line 106 on page 103, at first, the distance (rest) is calculated as the sum of the squares of the differences between a point on the first contour and a point on the second contour, over the upper and lower contours, where the top contour difference is weighted by the topToBottom variable. This distance (rest) is used in subsequent iterations to determine the horizontal, vertical and diagonal difference values in the loop beginning at line 137 on page 103. To determine each of these values, the current distance value, represented by rest, would be added to the previous values in the down, left, and down-left array positions, the down-left position value being the diagonal position which is weighted by the centerWeight factor as previously described. Referring to FIG. 26A, which illustrates the positional relationship between a previously determined value X, at array location 502, and subsequent array locations, the value X might be added to the difference values of subsequent locations to accumulate the total difference. calculations is shown. When calculating the difference value for array location 504, the value in location 502 would be used as the down value. Similarly, when calculating the value in location 506, the value of location 502 would be used as the center-weighted down-left, or diagonal, value. After calculating the three difference values, steps 458, 460, and 462, the process continues by selecting the smallest of the three values, step 464, for insertion into the current array position, step 466. As illustrated in the Appendix at line 144 of page 103, the FMin() function from page 101 returns the minimum of the three values previously calculated, the value being inserted into the storage array pointed to by pointer dc.

Subsequently, the process illustrated in FIG. 25 continues by determining the differences between the point on the first contour, represented by L1, to points on the second contour, represented by L2. Decision step 468 controls the iterative processing of the points along the second contour by testing for the end of the contour, or swath. In the implementation shown in the Appendix, the index variables i and j are used in place of L1 and L2 to control the difference calculation loops. As indicated in the code for the NewMatch function beginning on page 102 of the Appendix, the swath is referred to as the bandwidth, and is determined by a desired bandwidth which is adjusted for the slope defined by the contour lengths (see page 102, lines 83-89). If no limit has been reached, processing for the next point would continue at step 458 after the value of L2 was incremented at step 470. Similarly, decision step 472 controls the processing of each point along the first contour, in conjunction with incrementing step 474. Once all the points have been processed with respect to one another, as evidenced by an affirmative response in step 472, the relative difference score, best score, is contained in the farthest diagonal position of the array (L1, L2). Subsequently, the value determined at step 476 is returned as an indication of the dynamically warped difference between the contours being compared.

The code implementation found in the NewMatch() function on page 103 of the Appendix has optimized the execution of the aforedescribed warping process by reducing the large two-dimensional array to a pair of linear arrays which are updated as necessary. Due to this modification, the minimum difference, or best score, for the warp comparison value is found in the last location of the one-dimensional array. Furthermore, the final difference value, dc, may be subsequently normalized to account for the length differences between the two sets of contours being compared. Finally, such a value might subsequently be compared against a threshold or a set of similarly obtained difference values to determine whether the contours are close enough to declare a match between the words, or to determine the best match from a series of word shape comparisons.

In yet another embodiment, the dynamic time warping process previously described may be altered to compare the difference values contained in the difference array to a threshold value on a periodic basis. Upon comparison, the process may be discontinued when it is determined that sufficient difference exists to determine that the contours being compared do not match one another, possibly saving valuable processing time. Moreover, the sequential operation of word shape comparator 26 might be done in conjunction with sequential output from word shape computer 24, thereby enabling the parallel processing of a textual image when searching for a keyword.

Having described a basic implementation of the dynamic warping comparison measures, the distinctions of the other dynamic warp comparison methods included in the Appendix and the application of the control factors previously mentioned will be briefly described to illustrate the numerous possible embodiments of the present invention. First, the method previously described may also be implemented with the slope of the warp path being constrained as it moves across the array. Details of the implementation are found in the SlopeCMatch() function beginning on page 111 of the Appendix. This measure is further illustrated graphically in FIG. 26B, where the value of array location 512, X, may be added to only the three subsequent array locations shown. For example, X may be added to array location 514, when considered as the d2L1 value for location 514. The nomenclature used for the variable names, and followed in the figure, is as follows: d2L1 refers to the array location which is down 2 rows and left one column, dlL1, refers to the lower left diagonal array location, and dlL2 refers to the array location that is down one column on left 2 rows from the current array location. In a similar manner, X may be added as the dlL2 value for the calculation of the cumulative difference value for array location 516.

As is apparent from a comparison of FIGS. 26A and 26B, the slope constrained warping measure limits the warping path which can be followed during the generation of the cumulative difference value. The reason for implementing such a constraint is to prevent the warping process from removing, or compressing, a large area of one of the two contours being compared, without imposing a significant "cost" to such a compression.

Next, the method previously described with respect to the parallel warping process may also be implemented on only one pair of contours at a time, for example, the upper contours of two word shapes. The functions SepMatch() and SepCMatch(), as found in the Appendix on pages 104 and 113, respectively, implement the separate matching measure in both the non-slope-constrained and slope-constrained fashions previously described. In general, these measures separately calculate the difference between the top or bottom contours of a pair of wordshapes. The general implementation indicated for the measures in the code shows that these measures are typically used sequentially, first determining the warped difference for the top contours, and then adding to it the warped difference from the bottom contour comparison, resulting in a total difference for the wordshapes.

By carrying out the comparison methods described in a "piece-wise" cascaded fashion, further processing benefits may also be derived. More specifically, cascaded comparison would entail, first, utilizing the upper contours of the words being compared to identify a word, or at least narrow the set of possible alternatives and, second, using the lower contour comparison to provide complete identification. It is believed that such an approach to word shape comparison operation 26 would considerably reduce processing time spent on identifying unknown word shapes by comparison to a dictionary of known word shapes, 28, as illustrated in FIG. 3B. Important to the cascaded comparison, is the constraint that the top and bottom warps applied to the contours must be relatively equivalent. This requirement arises from the fact that the upper and lower curves have a relationship to a common word, and if this relationship is not maintained during the warp analysis, the accuracy of the comparison will be compromised.

Alternatively, the dynamic warping technique may be applied as described, with the addition of a function suitable for accumulating the relative warp applied to the upper and lower curves in achieving the best match. For example, when a known, non-italicized word shape is compared to an unknown word shape, a shift in the warp applied to the upper curve relative to the lower curve could be indicative of an italicized word, however, the length of the warped region will remain the same for the top and bottom warps. Such a technique may prove useful in the identification of important words within a larger body of text, as these words are occasionally italicized for emphasis.

One of the control factors which has not been previously described is the bandwidth factor. As implemented, the bandwidth factor controls the relative width of the signal band in which the warping signal will be constrained. More specifically, the band width limitation is implemented by defining a region about the array diagonal in which the warp path which traverses the array is constrained. The constraint is implemented by assigning large values to those areas outside of the band width, so as to make it highly unlikely that the path would exceed the constraint.

Another factor which was briefly mentioned is the topToBottom factor. When applied, the value of this variable is used to weight the difference value determined for the top contour warping process. Therefore, use of a number greater than one, will cause the upper contour difference to be weighted more heavily than the lower contour difference. A very large number would effectively eliminate the lower contour difference completely and, likewise, a zero value would eliminate the upper contour difference completely. This factor is generally considered important to enable the upper contour to be weighted in proportion to its information content, as it generally carries more information regarding the word than does the lower contour.

The hillToValley ratio is a variable which is usually applied in situations when a known, or model, set of word shape contours is being compared against a set of word shape contours from an unknown image. In exercising this option, the model set of contours is passed as the comparison measure functions, for example, NewMatch() on page 102 of the Appendix. When determining the difference between points on the contours, the comparison functions commonly call the function SquareDifference() on page 101 of the Appendix to determine the sum of the squared difference. SquareDifference() applies the hillToValley ratio to the squared difference whenever it determines that the value of the model contour is less than the contour being compared. The result of applying a hillToValley value greater than one is that the relative "cost" of the difference when the model contour is less than the target contour is smaller than the same difference when the model contour is greater than the target contour. The basis for this type of weighing is that when comparing against a model contour, the comparison should treat those areas of the target contour that are subject to being " filled in" during a scanning or similar digitizing operation with less weight than regions not likely to be filled in, as evidenced by contour positions below the model contour. For instance, the regions where ascenders and descenders meet the body of the character are likely to be filled in during scanning, thereby causing the target contour to have a gradual contour in those regions, whereas the model contour would most likely have a defined peak or valley in these regions. Hence, the contour value of the model would be less than the contour value of the target, even though the characters may have been identical. Therefore, the hillToValley variable attempts to minimize the impact to the calculated difference value over these regions.

It is important to note that the aforedescribed measures and control factors allow the comparison measures to be conducted in numerous permutations. However, the flexibility which these measures permit is intended to enhance the applicability of the comparison process, so that when information is known about a particular word shape contour, for example, a model contour generated from a computer generated character font, the measures may place reliance on that information to make the comparisons more robust.

The mathematical explanation of the word shape derivation process suggests that alternative methods of deriving the word shape signal exist. Some possible alternatives are the establishment of the one dimensional signal using an alternative coordinate scheme, for example polar coordinates. Another possibility is generation of signal g(t), where g(t) represents the direction from each contour point to the succeeding contour point, where t would represent the point number.

Depending on the particular application, and the relative importance of processing speed versus accuracy, for example, comparisons of different degrees of precision can be performed. For example, useful comparisons can be based on length, width (height) or some other measurement dimension of the image unit (or derived image unit shape representation, e.g., the largest figure in a document image); the location of the image unit in the document (including any selected figure or paragraph of a document image, e.g., headings, initial figures, one or more paragraphs or figures), font, typeface, cross-section being a sequence of pixels of similar state in an image unit); the number of ascenders; the number of descenders; the average pixel density; the length of a top line contour, including peaks and troughs; the length of a base contour, including peaks and troughs; and combinations of such classifiers. It has been found that an adequate comparison for purposes of determining phrase frequency is to compare only the length and height of the derived image unit shape representations. Such a comparison is particularly fast, resulting in a highly efficient phrase frequency analysis which has proven to be sufficiently robust to reliably extract significant phrases in many text document applications.

In instances in which multiple page documents are processed, each page is processed and the data held in the memory 15 (see FIG. 2), as described above. The entirety of the data can then be processed.

One way in which the image units can be conveniently compared and classified into equivalence classes is by comparing each image unit or image unit shape representation when it is formed with previously processed image units/shape representations, and if a match is obtained, the associated image unit is identified with the matching equivalence class. This can be done, for example, by providing a signal indicating a match and incrementing a counter or a register associated with the matching equivalence class. If the present image unit does not match with any previously processed image unit, then a new equivalence class is created for the present image unit.

Alternatively, as shown (step 50) the image units in each equivalence class can be linked together, and mapped to an equivalence class label that is determined for each equivalence class. The size of each equivalence class is indicative of the frequency with which the image units in the class occur in the document image.

Next (step 60), the sequences of equivalence class labels corresponding to all of the possible sequences of mapped image units in reading order (phrases) are analyzed to determine the frequency with which particular equivalence class label sequences, and hence the frequency with which the corresponding word unit phrases occur in the document. As used herein, a phrase is any sequence of adjacent word units which have been evaluated in reading order in the document. (It is not necessary that the stop or function words be discriminated prior to conducting the phrase frequency analysis.) Thus, the phrases include a master phrase consisting of the entire sequence of word units contained within a document image (or a portion thereof which has been selected for analysis), plus all of the subphrases contained within the master phrase.

Any conventional method for sorting symbol strings or sequences and determining recurring subsequences can be used to evaluate the equivalence label sequences. Exemplary methods are described in Knuth, The Art of Computer Programming, Sorting and Searching, (Addison-Wesley, 1973), Vol. 3 which is incorporated herein by reference. Advantageously, since the frequency of word units in a phrase will have a frequency at least as great as the frequency of the phrase, a useful preliminary sort of the equivalence label sequences is to eliminate any sequences containing word units which occur only once or twice. Similarly, an initial sort advantageously searches for sequences in which each of the word units in the sequence is a high frequency word.

The determination of what frequent phrases to identify as semantically significant phrases for purposes of further document processing, such as, for example, document summarization or categorization, depends on such variables as, for example, the length and style of the document, as well as the genre of its subject matter (e.g., technical literature versus fiction). Depending on all of the variables present in a particular situation, it may be advantageous either to select shorter length phrases having a relatively high frequency of occurrence, or longer length phrases having a lower frequency of occurrence. For example, in technical documents, it has been determined that two- or three-word phrases that occur at least about five times tend to be technical phrases indicative of the subject matter of the document. It will be appreciated that the choice of the most advantageous combination of frequent phrase selection parameters advantageously is determined empirically for a given application, and that the selection parameters are made settable so that they can be tuned for different applications.

A salient feature provided by the technique of the invention is the processing, identification, comparison, or manipulation of image unit sequences without an accompanying requirement that the content of the image units be decoded, even for output. More particularly, phrases are determined, processed and delivered for output without decoding, so that in essence, the actual content of the image units or recognition of the phrases is never required to be determined. Thus, for example, in such applications as copier machines or electronic printers that can print or reproduce images directly from one document to another without regard to ASCII or other encoding/decoding requirements, image unit phrases can be identified, and processed using one or more morphological image characteristic of the image units. In the comparison process described, for instance, selected image units, of undetermined content, in the area of the document image of interest are compared with other image units in the document also of undetermined content. Entire phrases, still of undetermined content, can then be optically or electronically delivered for output, for example, to an image reproducing apparatus of a copy machine, an electronic memory, a visual display, or the like, for example in producing a list of significant phrases in order of frequency of appearance in the document image.

The technique described above can be used to determine the significance of the phrases of a document, based upon the criterion of frequency of occurrence of a particular phrase. Thus, for example, the number of times a phrase appears can be used to construct a hierarchy of phrases, such hierarchy being useful for many purposes. It is noted, however, that the classifiers are determined without actually decoding the content of the image units; only the selected image classifiers of the image unit itself are used. The method can be applied, of course, to documents of multiple page length in a similar manner to that described above.

Although the invention has been described and illustrated with a certain degree of particularity, it is understood that the present disclosure has been made only by way of example, and that numerous changes in the combination and arrangement of parts or steps can be resorted to by those skilled in the art without departing from the spirit and scope of the invention, as hereinafter claimed. ##SPC1##