US 5,559,902 AGrant
Method for Enhancing Connected and Degraded Text Recognition
Issue Date:1996-09-24
•6 Claims
•11 Drawing Sheets
Abstract
The present invention provides a method and apparatus for enhancing and recognizing connected and degraded text. The enhancement process comprises filtering a scanned image to determine whether a binary image value of an image pixel should be complemented, determining whether complementing the value of the pixel reduces the sharpness of wedge-like figures in the image, and complementing the binary value of the pixel when doing so does not reduce sharpness. The recognition process may comprise determining primitive strokes in a scanned image, segmenting the scanned image into sub-character segments based on the primitive strokes, identifying features which characterize the sub-character segments, and comparing identified features to stochastic models of known characters and determining an optimum sequence of known characters based on the comparisons through the use of Viterbi scoring and level building procedures.
Metadata
Assignee
- Lucent Technologies Inc.
Inventors
- Chinmoy B. Bose
- Shyh-Shiaw Kuo
Application Information
Application Number:US 2516764
Filing Date:1994-05-31
Priority Date:1991-12-23
Art Unit:263
Classifications
IPC:
G06K 940
Field of Search:
38235834854;33;47;257;258;259;262;263;269;275463625;627
Patent Drawings (11 sheets)
Description
FIELD OF THE INVENTION
This invention relates generally to the field of optical text recognition, and specifically to the recognition of connected and degraded text.
BACKGROUND OF THE INVENTION
In modem business office environments, many devices and systems are used to improve the speed and efficiency associated with the creation, processing, and dissemination of documents. Among these are text processing systems, fax machines, and photocopiers.
From time to time, it may be necessary to convert the text of a printed document to electronic form for text processing or communication purposes. Such a circumstance may arise, e.g., when a document created on one text processing system must be edited on another system with which there exists no electronic communication capability. The conversion process for such text may comprise optical scanning and image analysis processes. The aim of the conversion process is the generation of a computer text file, typically comprising ASCII characters, which reflects the printed text. If a printed document comprises clean, well-formed text, this conversion process may not present much difficulty.
Because of distortion effects associated with repeated photocopying and facsimile transmission, certain documents may include fuzzy, swollen (degraded) and overlapped (connected) characters which make the text conversion process problematic. The greater the degree of degradation and connectivity, the more difficult it is to accurately discern and identify printed text characters. Naturally, computer files which result from the conversion of documents which contain such text frequently include errors in their representation of the documents words and characters.
SUMMARY OF THE INVENTION
The present invention provides a method and apparatus for recognizing connected and degraded text. An image preprocessing technique is provided which reduces the amount of spurious noise in scanned text images. Also provided is a segmentation process by which images are divided into a series of sub-character segments. A feature extraction procedure represents sub-character segments in terms of a set of predefined features. Identified features of segments are compared to stochastic models of known characters and character sequences for the purpose of text recognition.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 presents an illustrative text recognition process according to the present invention.
FIG. 2 presents line adjacency graph and compressed line adjacency graph representations of the character X.
FIG. 3 presents line adjacency graph and compressed line adjacency graph representations of the character e.
FIG. 4 presents the illustrative word pre-processing process presented in FIG. 1.
FIG. 5 presents a 3.times.3 window of pixels used in a nominal filtering process of a modified median filter.
FIG. 6 presents an illustrative original noisy image of the character w.
FIG. 7 presents a 3.times.7 window of pixels used in the modified median filter to preserve the sharpness of V-like shapes in an image.
FIG. 8 presents a 3.times.7 window of pixels used in the modified median filter to preserve the sharpness of inverse V-like shapes in an image.
FIG. 9 presents the character w having two noise-like pixels to be removed by a line adjacency graph filter process.
FIG. 10 presents the result of word preprocessing on the character image presented in FIG. 6.
FIG. 11 presents a line adjacency graph for the character e and associated strokes representative thereof.
FIG. 12 presents a line adjacency graph for the character x and associated strokes representative thereof.
FIG. 13 presents the set of strokes associated with the word hello.
FIG. 14 presents two adjacent strokes and the quantities related thereto used to determine whether such strokes should be merged.
FIG. 15 presents an illustrative stroke and an illustrative arc.
FIG. 16 presents several segments of a line adjacency graph for the characters ky.
FIG. 17 presents a set of features extracted from un-preprocessed segments associated with the characters ky.
FIG. 18 presents the features extracted from preprocessed segments of the characters ky.
FIG. 19 presents a line adjacency graph for the character z and associated strokes representative thereof.
FIG. 20 presents a set of primitive features extracted from segments associated with the word hello.
FIG. 21 presents the line adjacency graphs, compressed line adjacency graphs, and associated primitive feature strokes for the character i.
FIG. 22 presents an illustrative collection of 32 feature centers used by the illustrative embodiment of the present invention.
FIG. 23 presents the Hidden Markov Models for the characters ju.
FIG. 24 presents a trellis representative of the Viterbi scoring and level building techniques.
FIG. 25 presents illustrative overlap and blur parameter data for use in generating a training data set.
DETAILED DESCRIPTION
A. Introduction
FIG. 1 presents an illustrative text recognition process 10 according to the present invention. Process 10, which receives binary pixel images of individual words to be recognized, comprises a word preprocessing process 100, a subcharacter segmentation process 200, a feature extraction process 300, a training process 400, and a recognition process 500. The images received by process 10 may comprise connected (e.g., touching or overlapping) and degraded (i.e., noisy) characters. They are provided by text scanning and page preprocessing systems, 1 and 5 respectively. These systems 1,5 scan text from paper copies of documents, identify columns of printed text from the scanned images, identify lines within a column, and word boundaries within a line. Text scanning and page preprocessing systems known in the an may be employed for these purposes. See, e.g., H. S. Baird, Global-to-local layout analysis, Proc. IAPR Workshop on Syntactic and Structural Pattern Recog., (September 1988); and S. N. Srihari and G. W. Zack, Document Image Analysis, Proc. 8th Int'l Conf. Pattern Recognition, 434-436 (October 1986). In addition to providing images of scanned words, these systems provide estimates of character point size and base-line location.
Word preprocessing 100 performs filtering and other processing based on line adjacency graphs to reduce noise and retain word image sharpness. Sub-character segmentation 200 divides a preprocessed word image into a number of sub-character segments. These segments are defined using line adjacency graphs to identify strokes. Segments are defined based on the identified strokes. What results is a partitioning of the filtered pixel map received from word preprocessing 100 into a plurality of individual segment maps.
Following segmentation, feature extraction 300 is performed. Through feature extraction 300, each identified segment is characterized by one or more features which may be of either the stroke or arc variety. (If a word presented for recognition is constrained not to comprise connected or significantly degraded characters, it is possible to perform recognition based on comparing extracted features to feature models of known letters.)
With segments characterized by their features, process 10 may perform either of two processes: training 400 or recognition 500. By the training process 400, a Hidden Markov Model (HMM) is built for each text character to be recognized. Data associated with one or more trained HMMs (e.g., state transition probabilities) may be storm in a semiconductor memory (not shown), such as a Read Only Memory (ROM). Through the recognition process 500, stochastic distances of the sequences of unknown character segments are obtained based on the HMM state transition and associated bi-gram probabilities stored in memory. These distances are used to determine the most likely sequence of text characters which might have produced the unknown observed image segments. The most likely sequence of text characters may be saved in memory (not shown) for later retrieval.
Embodiments of the present invention may be used to augment the capabilities of conventional optical character recognition systems commercially available. This may be done by providing such systems with software performing the functions described below. Conventional systems would be required to perform the text scanning and page preprocessing tasks described above.
An illustrative set of software programs for an embodiment of the present invention written in the "C" language is provided in an Appendix attached hereto. The Appendix also provides a list of the programs associated with each of the word preprocessing 100, sub-character segmentation 200, feature extraction 300, training 400, and recognition 500 processes. These programs may be executed on computer marketed under the trademark SUN SPARCstation 1.
For clarity of explanation, the illustrative text recognition process 10 of the present invention is presented as comprising individual functional blocks. These functional blocks may be provided through the use of either shared or dedicated hardware, including, but not limited to, hardware capable of executing software. Illustrative embodiments may comprise digital signal processing (DSP) hardware, such as the AT&T DSP16 or DSP32C, and software performing the operations discussed below. Very large scale integration (VLSI) hardware embodiments of the present invention, as well as hybrid DSP/VLSI embodiments, may also be provided.
1. Line Adjacency Graphs
Line adjacency graphs (LAGs) are employed by several aspects of the illustrative process 10. As used in process 10, a LAG is stored in memory and represents run-lengths in a scanned and digitized image. Each "run" of consecutive black pixels on a scan line is denoted as a node of a LAG. The degree of a node is expressed as an ordered pair of numbers denoted as (a, b ). The number a equals the number of nodes above and connected to a given node, while the number b equals the number of nodes below and connected to the node.
A junction is a node of a LAG having a or b greater then one. A path is a node having a and b less than or equal to one. The left-hand portions of FIGS. 2 and 3 are LAG representations of characters X and e, respectively. In these figures, paths and junctions are indicated by solid and dotted lines, respectively.
The LAGs presented in FIGS. 2 and 3 can also be represented in a compressed form referred to as a c-LAG. In a c-LAG, connected paths can be represented in a compressed form referred to as a c-path. The right-hand portions of FIGS. 2 and 3 present c-LAG representations of characters X and e, respectively. Junctions are represented as circles, while c-paths are represented by shaded circles.
In case that one of the degrees of a junction is one, the junction is also included in the corresponding c-path connected to the junction if it is not an outlier compared to the nodes in the c-path. A junction may be considered to be an outlier if the width of the junction divided by the average width of the c-path exceeds a threshold, e.g., 1.4. For example, the two junctions of X in FIG. 2 are included into the corresponding c-path. However, the junction in the middle part of e in FIG. 3 is not included in the c-path connected to it, since it is an outlier.
B. Word Preprocessing
Illustrative word preprocessing 100 is performed on a presented word image in order to reduce spurious noise prior to training or recognition. As shown in FIG. 4, word preprocessing 100 comprises a modified median filter 120 and a LAG process 140. The modified median filter 120 reduces noise and preserves aspects of sharpness and connectivity while the LAG process 140 removes noise-like run-lengths in the image.
Modified median filter 120 comprises a nominal filtering process which is modified under certain circumstances. The nominal filtering process employs a 3.times.3 window of pixels, such as that shown in FIG. 5. The nominal process centers the window (window element no. five) over a given pixel in the image and assigns to that pixel the binary value associated with the majority of the pixels in the window (i.e., the binary value held by at least five of the nine pixels).
Two rules modify this nominal filtering process. The first rule concerns a situation when, according to the nominal filtering process, an empty (i.e., white) pixel should be filled (i.e., made black), such as pixel (i, j) in FIG. 6 (where i and j denote a specific row and column of the image). A 3.times.7 window, illustratively presented in FIG. 7, is centered over pixel (i, j). If more than 14 of the shaded pixels are filled (i.e., black) and both pixels (i-1, j) and (i-2, j) are empty, then the pixel (i, j) shall not be filled. This first rule preserves the sharpness of wedge-like shapes in the image which can enhance performance in subsequent processing steps. In this illustrative procedure, the wedge-like shapes preserved are V-like.
A like procedure may be performed using the 3.times.7 window of FIG. 8 to preserve the sharpness of inverse V-like wedge shapes. In this procedure, if more than 14 of the shaded pixels are filled and both pixels (i+1, j) and (i+2, j) are empty, then the pixel. (i, j) shall not be filled.
The second rule for modifying the nominal filtering process concerns the situation when, according to the nominal process, a filled pixel should be emptied. If such a pixel is an element of a sequence (or run-length) of at least five consecutive filled pixels, then the pixel is not emptied. This second rule may preserve connectivity of run-lengths which may be broken by the nominal filtering process of median filter 120.
Word preprocessing 100 further comprises a LAG filter process 140 to remove some noise-like run-lengths by determining and checking the LAGs associated with the image. Every path (i) located at the top or bottom of each blob having degree (0,1) or (1,0), respectively, and (ii) connected to a junction is removed, such as those two paths located at the top left of `w` in FIG. 9 (where a blob is any set of one or more pixels, wherein each pixel in the set is connected to at least one other pixel in the set in any of the eight ways one pixel may be connected to an adjacent neighbor pixel (vertically: up, down; horizontally: left, right; and diagonally: up-right, up-left, down-right, and down-left)). See Appendix, modules prep.c, lag.c, and clag.c.
FIG. 10 presents the character image w from FIG. 6 after operation of word preprocessing 100.
C. Sub-Character Segmentation
The sub-character segmentation process 200 divides image information received from the word preprocessing 100 into segments which can, in turn, be used to characterize the image in terms of segment features. Such characterization is useful to both the training and recognition processes, 400 and 500, respectively. Segmentation is carded out by identifying the strokes present in an image.
1. Stroke Identification
In the illustrative process 200, stroke identification is performed by first determining the direction of dominant strokes in the image. Dominant strokes may be identified by scanning the pixel profiles within a preset range of angular directions at small intervals, and choosing the direction of tallest peaks in the profile. If the direction of the dominant strokes is not vertical, the direction may be normalized by rotating the pixels on an imaginary slanted line (based on the slant angle) such that the pixels fall on a vertical line.
Primitive strokes are identified to provide useful structural information for segmentation. The primitive strokes in a word can be identified by generating and analyzing a c-LAG of the image to be recognized. See Appendix, modules lag.c and clag.c. A primitive stroke is identified by its endpoints in the two-dimensional plane in which it lies, x.sub.1, Y.sub.1 and x.sub.2,Y.sub.2, where values for x and y are related to the top left comer pixel of a rectangle surrounding a scanned blob. The rectangle has dimensions equal to the height and width of the blob. See Appendix, module blob.sub.-- extr.c.
Each c-path of the c-LAG is analyzed according to its own characteristics and its neighborhood information. The first step of analyzing a c-path is to divide it into one or more groups of nodes which have similar width and collinear centers.
Consecutive nodes (indicated by i and i+1) will be considered to have dissimilar widths, w(i) and w(i+1), if all the following tests are satisfied: ##EQU1## wherein, e.g., .alpha.=2.0, .beta.=0.7, and .differential.=0.15. For a group which comprises nodes of similar widths, collinearity of node centers is determined by defining a line through the centers of the first and last nodes in the group and determining the maximum distance of any node center in the group from this line. If this maximum distance is, e.g. less than 2.6 pixel units, the nodes of the group are said to have collinear centers.
Strokes corresponding to each group with nodes of similar widths and collinear centers are identified (or returned) according to the following rules:
i. When the ratio of group height over average width of the group, denoted as R.sub.h/w, is larger than a threshold (e.g., 1.50), a vertical stroke is returned which is a line fitting centers of the nodes.
ii. When the ratio R.sub.h/w is smaller than a threshold (e.g., 0.65), a horizontal stroke is returned which lies in the middle of the group.
iii. When a group is adjacent to a much wider junction or path of another group either at top or bottom, a vertical stroke is returned (see FIG. 11 ). A group is said to be much wider than another if its width at the point of adjacency divided by the average width of the other group is greater than a threshold, e.g., 1.7.
iv. If a c-path contains only one group and that group connects to two c-paths at
both top and bottom, two crossed strokes are returned (see FIG. 12).
Each vertical stroke is characterized by a width which is used in segmentation. The width of a vertical stroke is defined as the average width of nodes in the group from which the stroke is returned. Horizontal strokes, on the other hand, are not characterized by widths since such width information is not used by the segmentation process 200.
In order to avoid ambiguity in segmentation process 200, strokes are not returned from ambiguous c-paths, i.e., those c-paths which do not satisfy any of the rules (i-iv) for returning strokes. For example, in FIG. 13, no stroke is returned for the part of the image where "lo" touches and the upper right part of "o."
The final step for stroke identification in the illustrative segmentation process 200 is the merging of certain adjacent strokes. Strokes may be merged if deviation to be incurred as a result of the merge is within a predetermined tolerance. Consider the example in FIG. 14, where E.sub.1, E.sub.2, E.sub.3 and E.sub.4 are the endpoints of the two adjacent strokes. A new merged stroke is formed by connecting the starting point of the first stroke, E.sub.1, to the end point of the second stroke, E.sub.4. Then, five quantities are checked:
three distances, E.sub.2 P.sub.2, E.sub.3 P.sub.3 and E.sub.2 E.sub.3, and two ratios, ##EQU2## If all the distances and ratios are smaller than predetermined thresholds (e.g., 2.2, 2.2, 5.1, respectively, for the distances, and 1/7.4 for both ratios), the deviation is deemed acceptable and the two original strokes may be replaced by a new merged stroke. The threshold values are functions of scanning resolution (pixels per unit length-assumed known), font type and size (assumed available from the page layout preprocessor), and may be set with a look-up table.
2. Segmentation Rules
Sub-character segmentation is achieved by applying a set of rules based on the returned strokes. The segment boundaries obtained by applying these rules partition the original pixel image into individual image segments. The segmentation rules are as follows:
i. A non-horizontal stroke without any vertical overlap with any other stroke identifies a non-horizontal segment, where vertical overlap refers to one stroke partially or wholly lying above or below another stroke as viewed from a vertical direction. The width of the segment is obtained from the width of its strokes.
ii. The space between two non-horizontal segments identifies a horizontal segment.
iii. The vertical overlap of two vertical (or near-vertical) strokes or two inclined strokes identifies a non-horizontal segment with a width determined by the overlapped width of the individual strokes. Specifically, non-horizontal segment width refers to the lateral distance traversed by one or more vertically overlapping non-horizontal strokes plus an additional distance added to each stroke end. This additional distance is a fraction of the average path width of the paths forming the strokes. This fraction depends on the angle which the stroke makes with the vertical. See Appendix, module blob2feat.c. Because there may be more than one stroke in a segment, different non-horizontal strokes may define the left and right edges of a segment. Consequently, the average path width added to define each edge may not be the same.
iv. The vertical overlap of a vertical stroke with any other non-vertical stroke provides the segment boundaries dictated by the vertical stroke.
v. The vertical overlap of an inclined stroke with a horizontal stroke provides the segment boundaries dictated by the inclined stroke.
vi. Two intercepting inclined strokes with slopes of opposite sign (e.g., strokes forming an `x` pattern) provide a segment boundary at the point of intersection.
D. Feature Extraction
Once a pixel image is segmented by the segmentation process 200, the individual segments may be characterized by identifying in such segments one or more features. Such identified features may be used both in training 400 and character recognition 500.
In the illustrative feature extraction process 300, two types of features are identified within segments: strokes and arcs. FIG. 15 presents an illustrative stroke and an illustrative are. A stroke, which is a line-segment, is uniquely identified by its centroid, length, and slope, and may be represented by a 5-tuple (x, y, rsin2.theta., rcos2.theta., d), where (x, y) is its centroid measured with respect to base-line information provided by systems 1,5, r is its length and .theta. is the slope angle. (Twice the slope angle is used for the purpose of maintaining continuity in the parametric representation, as the slope angle varies between -90 and +90 degrees.) The value d is always 0 for a stroke. An are may also be represented as the 5-tuple (x, y, rsin 2.theta., rcos2.theta., d), where the first four parameters represent a chord of the arc (in the same fashion as a stroke), and d is the maximum perpendicular distance of the are from the chord.
1. Segment Preprocessing
Prior to feature identification, the illustrative feature extraction process 300 preprocesses individual segments to remove certain noise-like pixels. For example, FIG. 16 presents several segments of the character string ky. If the illustrative feature identification technique is applied directly to these segments, some undesired features will be extracted due to groups of noise-like pixels as indicated in FIG. 17 by labels a-f.
Consider the group of noise-like pixels identified by the label e. Since this group is actually a small portion of a c-path of the letter y, it can be excluded from the second segment (which concerns a portion of the letter k). Exclusion of noise-like pixels is done by eliminating all pixel groups which are a portion of either (i) a path or (ii) a c-path in a neighboring segment. FIG. 18 presents the features which are extracted from preprocessed segments of string ky. Strokes associated with noise-like pixel groups a-f are no longer present.
2. Feature Identification
A structural analysis similar to that described above for sub-character segmentation is employed for identifying (or extracting) segment features. The first step in this analysis involves representing each image segment by a c-LAG. Each c-path of a c-LAG is then analyzed to identify its features. If a segment is identified as horizontal, horizontal strokes are returned from each c-path of the segment. See Appendix, path.sub.13 s.c. For non-horizontal segments, each c-path thereof is checked and subdivided into groups of nodes, if applicable.
The process of subdividing c-paths for feature identification is different from that performed for sub-character segmentation (where a c-path is subdivided based upon either a large width change between two adjacent nodes or non-collinear node centers). Here, groups are formed by checking for width change only as described above.
Two adjacent groups in a segment will be merged into a single group if the following two conditions are satisfied: ##EQU3## Where w.sub.1 and w.sub.2 denote the average widths of the two adjacent potential groups, and .alpha. and .beta. are predetermined constants (e.g., 3.0 and 0.6, respectively).
The purpose of conditional merging of groups is to preserve arc-features within segments. As shown in FIG. 18, them are two potential groups which might be identified in the first segment of the character y due to significant node width changes (at the bottom of the character). However, because the change of widths between the two potential adjacent groups is not large enough, as determined by the above conditions i and ii, the c-path contains only one group. Thus an arc is able to be extracted from that segment according to the criterion discussed below. In contrast, FIG. 19 shows a LAG for the letter z wherein the c-path beginning at the top of the letter is subdivided into two groups which cannot be merged under conditions i and ii. Therefore, separate corresponding strokes may be identified according to the rule discussed below.
Arc and stroke features are identified sequentially in each group of nodes within a segment. Arcs may be identified by constructing a line connecting the centers of the first and last nodes in a group. The center of a node, within the group, located at the greatest distance from the line is then determined. If the ratio of this largest distance over the length of the line is larger than a threshold (e.g., 0.1), an arc is identified and returned from the group. For example, in FIG. 20, an arc--indicated by a triangle--is returned from the 4th and 10th segments. The three vertices are the centers of the first and last nodes, and the node center located at the maximum distance from the line.
Generally, the same rules discussed above used in defining strokes for purposes of sub-character segmentation may be used here. As such, a stroke may be identified based on the ratio of height over average width of the group (R.sub.h/w). Unlike stroke definition for segmentation, however, looser thresholds may be used in order to return strokes from most of the groups (e.g., 1.2 and 0.85 are used instead of 1.5 and 0.65, respectively).
Special rules may be used for those unclear groups, such as:
i. For an isolated single c-path, such as, the top part of the character `i` in FIG. 21, return a vertical stroke if R.sub.h/w is larger than 0.9. Otherwise, return a horizontal stroke.
ii. If any vertical stroke identified during segmentation is contained in a segment, each ambiguous group within this segment returns a vertical stroke as a feature.
iii. If a junction is the first or last node in a c-LAG, a horizontal stroke is returned. See, e.g., the horizontal stroke at the bottom of character `z` in FIG. 19.
Each identified feature is represented as a 5-tuple in a continuous vector space. In both the recognition and training modes, these vectors are mapped to a discrete space defined by a clustering algorithm (see the section on Training).
E. Training
1. Introduction
Identifying unknown connected and degraded character images is accomplished by illustrative process 10 by relating observed features, extracted as described above, to known features of known characters. The "closer" observed features are to a stochastic model of a given known character, the more confidently the image which produced the features can be identified as the known character.
Depending on the appearance of characters in an image, features extracted from a given segment of a given character image may not always be the same (for any two samples of the character to be identified). Characters may appear differently, for example, due to varying connectivity with neighboring characters and varying character degradation (or blur). Furthermore, the starting and ending points of the individual characters become obscured.
Observing features of connected and degraded character images and determining the characters to which such features correspond depends upon a doubly embedded stochastic process. That is, one which has an underlying observable stochastic process concerning which features might be extracted from image segments, and another stochastic process, not directly observable, concerning which stochastic features might be associated with extracted features.
Illustrative process 10 represents the doubly embedded stochastic processes associated with connected and degraded text recognition through the use of Hidden Markov Models (HMM). Unlike the discrete observable Markov Model, wherein each model state corresponds to an observable event, the states of a HMM are not directly observable. Rather, observations are probabilistic functions of the state to be determined. A HMM is provided for each character to be recognized. See Appendix, modules recinit.c and ndorec.c. Each state of a HMM represents one segment of a character. Thus, the number of states in a model depends upon the number of segments needed to represent a character.
Each HMM of illustrative process 10, .lambda., may be described generally as follows:
i. Each model comprises a set of states:.OMEGA.={.omega..sub.j : 1.ltoreq.j.ltoreq.J}, where J is the number of states in the model. Each state is a stochastic representation of a segment of a character to be recognized.
ii. Each model has associated with it a matrix of state transition probabilities: A={a.sub.jm : 1.ltoreq.j,m.ltoreq.J}, where a.sub.jm =P(.omega..sub.m at i+1.vertline..omega..sub.j at i). These probabilities represent the likelihood that, for a given model, one state (or segment), .omega..sub.m, will follow a given state (or segment), .omega..sub.j, in time.
iii. For each state of a model, a vector of observation probabilities for observation X.sub.i : B={b.sub.j (X.sub.i)}, where b.sub.j (X.sub.i)=P(X.sub.i .vertline..omega..sub.j at i ). These probabilities represent the likelihood that a given observed segment vector, X i, is associated with a given state, .omega..sub.j (see section 4, below).
iv. Associated with each state of a model is an initial state probability: II={.pi..sub.j }, where .pi..sub.j =P(.omega..sub.j at i=1). These probabilities represent the likelihood that a given model state will be the initial state from which the first state transition will be made.
v. As part of the recognition process 500 discussed below, each state of each HMM is compared against each segment vector in an observation vector sequence:X={X.sub.i : 1.ltoreq.i.ltoreq.I}, where I is the number of observations. This vector represents the series of binary segment vectors representing the features extracted sequentially from the image of a character string.
In addition to the use of probabilities associated with each model of a character (i.e., state transition probabilities, a.sub.jm ; observation probabilities, b.sub.j (X.sub.i); and initial state probabilities, .pi..sub.j), illustrative process 10 employs measures of likelihood associated with the succession of characters in an image. Process 10 utilizes bi-gram probabilities to reflect the likelihood that one character will follow another in a word presented for recognition. Bi-gram probabilities provide contextual information to aid in the process of character and word recognition.
The training process 400 supplies the HMMs of the illustrative process 10 with information which may be used to analyze observations, X.sub.i, to determine a maximum likelihood solution to the problem of identifying connected and degraded characters. That is, training 400 provides not only the state transition probabilities, a.sub.jm, observation probabilities, b.sub.j (X.sub.i), and initial state probabilities, .pi..sub.j, but also the bi-gram probabilities for contextual analysis. Given an observation sequence X and the model parameters determined through training 400, a recognition process 500 may be employed to determine the optimal state sequence associated with the observations {.omega..sub.ji : 1.ltoreq.j.ltoreq.J, 1.ltoreq.i.ltoreq.I}. In other words, recognition 500 may determine the most likely sequence of characters which may be postulated given the set of observations.
2. Training Data Set
In order to have an appropriate training data set for deriving the HMM parameters, it is preferred that a character data set be generally representative of the expected characters in the words presented for recognition. A pseudo-random character generator of the type described by H. S. Baird, Document image defect models, Proc. IAPR Workshop on Syntactic and Structural Pattern Recog., (June 1990), may be used to obtain a set each of the characters for training. For example, the character generator may provide a training set comprising the lower case Roman alphabet (a-z), printed in Times Roman font (point size 10) and scanned (simulated) at 300 pixels per inch. The character generator should provide the two major sources of noise in printed text--overlap and blur. FIG. 25 presents illustrative overlap and blur parameter data for use with the character generator. In the figure, points indicated by an "S" indicate separate characters within a word while those indicated by a "T" indicate that characters are touching lightly. Overlap and blur are not orthogonal parameters; that is, a certain amount of blur may produce overlap. Nonetheless, it may be preferable to perform training with a certain amount of overlap not accomplished by blur parameters. A set of approximately 550 non-overlapping training characters at several blur levels may be generated by the character generator, keeping other parameters of the above-referenced character generator constant.
3. Clustering of Features
The training set of characters should be segmented and their features extracted, as described above, to produce, for example, a set of 1400 segments consisting of a total of approximately 2000 features. These features may be clustered using a k-means algorithm, such as that described by, A. K. Jain and R. C. Dubes, Algorithms for Clustering Data, Chapter 3 (1988), and J. A. Hartigan, Clustering Algorithms, Chapter 4 (1975). The algorithm may be started with a set of 15 visibly distinct duster centers chosen from one of the (a-z) training sets. A "compactness of cluster" index may be defined as: ##EQU4##
At the end of iterations of the clustering algorithm for the current number of centers, the mean of C.sub.c over all clusters should be determined (the distances in the numerator are weighted by the number of members). The number of clusters should be allowed to increase if the mean "compactness" increases, unless a predetermined number of clusters is reached. A new cluster center may be chosen for a new iteration as the most distant member of the worst (in the sense of the "compactness" criterion) cluster. For example, the algorithm may return a set of 32 cluster (or feature) centers from the set of approximately 2000 features. FIG. 22 presents an illustrative collection of 32 feature centers (the arcs are represented by triangles with the apex indicating the maximum deviation of the arc from the chord). Each dot in the field in which each feature center lies represents a comer of a pixel. The "+" represents the intersection of a vertical center-line and the character base-line. The feature centers are provided given 10-point scanned characters. They may be advantageously scaled based on input character point site information from systems 1,5. See Appendix, modules cluster.c and quant.c.
4. Vector Representation of Segments and Observation Probability
The clustering of features provides a way of partitioning the continuous feature space to a discrete feature space. A training character segment may be represented by a 32-bit binary segment vector, where each bit which is set identifies a given feature center of the discrete feature space which is closest to an identified feature in the segment.
Through the training process 400, observation probabilities are estimated with use of a Bayesian distortion measure using binary features, such as that described by R. O. Duda and P. E. Hart, Pattern Classification and Scene Analysis, Sec. 2 (1973), under the assumption of class-conditional statistical independence among the features. Training for each of the character models is performed by segmenting samples of each character, mapping features extracted from each segment to a binary segment vector, and associating with each feature extracted--each bit of the vector--a probability of occurrence (i.e., a probability that the feature associated with the bit position will be observed).
Each segment of a HMM may be labeled. For instance, the segments for the character `u` may be labeled as `u0`, `u1` and `u2`. These labels may also be used to represent the corresponding states of the HMM for each character.
Each state of a character model is characterized by the binary probability distribution associated with each bit location. If p.sub.n is an estimate of the probability density function at bit location n, where 1.ltoreq.n.ltoreq.N (e.g., N=32), then ##EQU5## where x.sub.n is the binary value of bit n in vector X={x.sub.n : 1.ltoreq.n.ltoreq.N}, .omega..sub.jk is the event of the state being j in model k, y jk (n) is the total number of times the bit n of the segment vector for state j, model k was set during training, and y.sub.jk is the total number of times state j of model k appeared during training.
Naturally, density function p.sub.n is but an estimate of the density, and would approach the real density when the sample size is large. Note that many of the bit probabilities in a character model may be zero after a training run, due to the difference in features corresponding to different states and different models. In order to resolve any computational problems which might result from this situation, a small probability may be assigned in place of all zero probabilities.
The observation probability for an observation X is: ##EQU6## assuming class-conditional statistical independence between features. Taking logarithm of the above expression (which simplifies the product to a sum, but retains the relative distance relationships), and redefining {b.sub.j (X)}, ##EQU7## This observation (log) probability serves as a Bayesian measure of distortion or distance of the observation vector X with respect to a model state.
5. State Transition Probability
Within a specific HMM for a character, the state transition probability is defined as:
where i is the observation sequence, 1.ltoreq.j,m.ltoreq.J and m.gtoreq.j. Given the physical ordering of the states within a model, a left-right sequence of HMM states is preferred. As such, the HMM for the character `u` where J=3 is:.omega..sub.1 =u0, .omega..sub.2 =u1, and .omega..sub.3 =u2. The state transition probabilities within a character may be estimated as: ##EQU8## where z.sub.j (m) is the total number of transitions from state .omega..sub.j to .omega..sub.m, and z.sub.j is the total number of transitions from state .omega..sub.j. Based on observations of connected characters, it is preferred to skip at most one state during the state transitions, i.e., m-j.ltoreq.2.
Meaningful transition probabilities between states of the same character model should be determined from a large representative training data set. For the illustrative training data set character models, the transition probabilities are highly dependent on the degree of connectedness (overlap), the connected pair of characters, and the mount of noise existing in the test samples. The degree of overlap and blur in the training data set should be representative of that expected for the system in operation so that transition probabilities will accurately reflect the likelihood of state (segment) succession in real character images to be identified.
Therefore, in lieu of the transition probabilities, penalty functions are added to the cumulative distance measures, to penalize for skipping a state or for staying in the same state. See Appendix, module ndorec.c. Performance may be enhanced by providing penalty functions which are tuned (or graded) for (i) different characters or (ii) important states in a given character model. The skipping of a state may be induced by the overlapping of two characters (for instance, overlap of the last segment in the segment string {j0, j1 }with the first segment in the segment string {u0, u1, u2}in the character string "..ju.." in FIG. 23). It could also be induced by a missing segment due to deformation of the character. The decision to stay in the same state may be caused by an extra segment generated because of deformation of the character.
The transition probability between character models--the bi-gram probabilities--may be determined from the statistical studies of the type of text material the recognizer is expected to handle. For general English text, the statistical results of previous studies performed on the transition probabilities between characters may be used, such as those provided by A. G. Konheim, Cryptography: A Primer, Sec. 2.3 (1981), which reports first order transition probabilities between two successive letters in the English language. See Appendix, module recinit.c. These probabilities are used in a level building technique described below. Although the illustrative process 10 employs bi-gram probabilities, an embodiment may employ n-gram probabilities (n>2) without requiting much in the way of additional computational burden.
6. Initial State Probabilities
Each character is represented by a HMM with its own initial probability which is assigned to the first and second states in a left-right model. (A second state initial probability is assigned to address the skipping of model's first state. ) Initial state probabilities apply to the model corresponding to the first character of a character string. This probability may be used at the start of the level building algorithm (described below) to discriminate between probable character strings. Again, the initial state probability may be estimated from the statistical studies of the type of text material the recognizer is expected to handle. In case of general English text, useful data is provided by A. Kundu, Y. He, and P. Bahl, Recognition of handwritten word: First and second order Hidden Markov Model based approach, Vol. 22, no. 3, Pattern Recognition (1989). This data is based on the dictionary entries of English words starting with each character.
F. Recognition
1. Introduction
For recognition of a character string, the segmentation technique for separating sub-character segments described above may be used. Using the process described above for the training mode, the binary feature vectors which correspond to each segment are found. The Bayesian distortion measure (defined above) for finding the distance of the observed segment from the statistical models for the trained segments is used. Modified Viterbi scoring is used to match the unknown connected segments against the single character HMMs. A level building procedure keeps track of the path yielding minimum distance (maximum probability) for the string up to any segment. Parallel processing techniques may be preferred for the recognition process 500 to minimize processing time. See Appendix, modules nrec.c and ndorec.c.
2. Viterbi Scoring
Let the states corresponding to I observations be defined as Q={q1, q2, .., qi, .., qI}. The best state sequence (that is, the one which maximizes P(Q.vertline.X), where X is the input observation sequence) is given by application of the Viterbi scoring procedure defined below: ##EQU9## where .delta..sub.i (j) is the best score (highest probability) along a single path at observation i, and .psi..sub.i (j) keeps track of the optimal states which provide such score. ##EQU10##
A trellis structure, such as that presented in FIG. 20 and described by S. E. Levinson, L. R. Rabiner, M. M. Sondhi, An introduction to the application of the theory of probabilistic functions of a Markov process to automatic speech recognition, Vol. 62, no. 4, Bell Syst. Tech. Journal, 1035-1074, (Apr. 1983), explains the implementation of Viterbi scoring (and the level-building technique). Each character is represented by a HMM, denoted as .lambda..sup.k, (1.ltoreq.k.ltoreq.K), where K is the total number of models, which may be greater than the number of characters to be identified. The HMMs allow state transitions (between two consecutive observations) either to the same state, or to the next state, or to the next higher state (skipping a state). The restrictions discussed above concerning penalty functions and state skipping are part of the model characteristics and reflect the nature of the character deformations. A character model is expected to start at the first or the second state. For each observation i, .delta..sub.i (m) is calculated for each state of each model. Based on the cumulative measure or score, a decision is made recursively on the optimum previous state at the previous observation (see FIG. 24). The above expressions for the termination and backtracking for a single model case have been modified in the level building algorithm described in the following section. For easier manipulations, the probabilities have been replaced by negative log-probabilities in the calculations of b.sub.m (X.sub.i) and .delta..sub.i (m ). (These are also referred to as "distances" herein.)
3. Level Building
In illustrative process 10 and recognition process 500, recognition is based on individual character models and is achieved through a determination of the optimum sequence of character models that best matches (in a maximum likelihood sense) the unknown character string (probably deformed and connected). The level building technique introduced above is applied to solve for such an optimum sequence of character models.
The level building algorithm is presented in FIG. 24. In this figure, i is an observation point corresponding to the observation X.sub.i, j is a state of the HMM .lambda..sup.k, l is a level of the stacked models (level corresponds to character position within the string), and k (axis perpendicular to the plane of the page) is the character corresponding to the model .lambda..sup.k. At each observation, the cumulative distances for each model and each state is updated for each level. (The operation may be trimmed considerably by noting that, based on the slope of the trellis and the maximum number of states among the HMMs, some areas of the trellis cannot be reached.) At the end of each level I for the same observation, a minimization of the cumulative distance is performed over all k for identifying the best model at that observation with respect to each succeeding model.
If the cumulative distance at the end of level I for observation i is defined as D.sub.l.sup.k (i), then the best model at observation i for the next character model p (d.sub.tr (k,p) is the transition probability from model k to model p) is as follows: ##EQU11## where C.sub.lp.sup.B (i ) stores the value of k corresponding to the best character model at observation i, level I and for the next character p. P.sub.lp .sup.B (i) stores the backpointer to the best model at the previous level corresponding to C.sub.lp.sup.B (i). i.sub.1 indicates the length (number of observations) of the current level for the model C.sub.lp.sup.B (i). Since skipping of a state within a model is allowed, D.sub.l.sup.k (i) holds the minimum of the cumulative distances at the last and next to last state of model .lambda..sup.k.
The initial best probability (shortest cumulative distance) at each new level I for each model p is obtained from the stored value of D.sub.lp.sup.B (i-1). The Viterbi score is incremented by matching the character models beginning at the new level. The best character string of length I may be identified at each observation i by backtracking the pointer P.sub.lp.sup.B (i) to l=1. This process continues recursively until the end of the maximum expected levels. The overall best string is obtained from min D.sub.lp.sup.B (I), where L is the maximum expected number of characters in the string, I 1.ltoreq.1.ltoreq.L is the last observation and d.sub.tr (k, p )=0. See Appendix, modules nrec.c and ndorec.c.
4. Context by Lexicon
In addition to the context provided by the use of bi-gram probabilities, context may be provided through the use of a dictionary or lexicon. Words which are identified by the Viterbi/level building techniques may be compared with a lexicon of words to see if such identified words are present. If not, the closest word in the lexicon may be used in place of the identified word or, the closest word may simply be noted for future use by an operator. ##SPC1##
Claims
We claim:
1. A method of generating an enhanced image based on a scanned image, the scanned image comprising image pixel signals having been generated with use of an image scanner, an image pixel signal having a binary value, the method comprising the automated steps of: filtering one or more image pixel signals of the scanned image to form a filter signal, the filter signal corresponding to a given image pixel signal; comparing a value of the filter signal with a value of the given image pixel signal; determining whether complementing the value of the given image pixel signal reduces the sharpness of a wedge-like shape in the image; and complementing the binary value of the given image pixel signal when (i) doing so does not reduce the sharpness of the wedge-like shape in the image and (ii) said filter signal value and given image pixel signal value are not equal, to generate an image pixel signal of the enhanced image.
2. The method of claim 1 wherein the step of performing a filter process comprises: centering a window of pixels on the given image pixel signal; and determining the value of the filter signal corresponding to the given image pixel signal based on the majority binary value of image pixel signals in the window.
3. The method of claim 1 wherein the step of determining whether complementing the value of the given image pixel signal reduces the sharpness of a wedge-like shape in the image comprises: centering a window of pixels on the given image pixel signal; determining whether a majority of image pixel signals which surround the given image pixel signal at the center of the window and which form a V-like wedge have a binary value opposite to that of the centered given image pixel signal; and determining whether a predetermined number of image pixel signals below the centered given image pixel signal each have a binary value equal to that of the centered given image pixel signal.
4. The method of claim 1 wherein the step of determining whether complementing the value of the given image pixel signal reduces the sharpness of a wedge-like shape in the image comprises: centering a window of pixels on the given image pixel signal; determining whether a majority of image pixel signals which surround the given image pixel signal at the center of the window and which form an inverse V-like wedge have a binary value opposite to that of the centered given image pixel signal; and determining whether a predetermined number of image pixel signals above the centered given image pixel signal each have a binary value equal to that of the centered given image pixel signal.
5. The method of claim 1 further comprising the step of removing noise-like run-lengths of image pixel signals located at the top or bottom of a scanned image.
6. The method of claim 1 wherein the step of complementing the binary value of the given image pixel signal comprises complementing the binary value when doing so enhances the sharpness of a wedge-like shape in the image.
Patent Citations (8)
Non-Patent Literature (22)
- Hoshino, T. et al., "Notch Elimination Method for Binary Figures Using Run-Length Coding," Systems, Computers, Controls, vol. 13, No. 1.
- Bozinovoc, R. M., "Off-Line Cursive Script Word Recognition," IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 11, No. 1, Jan. 1989, pp. 68-83.
- Vlontzos, J. A. et al., "A Hierarchical System for Character Recognition," IEEE Int'l Symp. on Circuits and Systems, 1989, May 1989, pp. 1-4.
- IBM Technical Disclosure Bulletin, vol. 34, No. 7B, Dec. 1991, "Two-Step System for Character Recognition Without Pre-Stored Fonts."
- European Search Report dated Sep. 1, 1994, Appln. No. 92311264.3.
- European Search Report dated Jun. 4, 1994, Appln. No. 92311265.0.
- H. S. Baird "Feature Identification for Hybrid Structural/Statistical Pattern Classification", Computer Vision, Graphics, and Image Processing, 42, 318-333 (1988), pp. 318-333.
- J. J. Hull "Hypothesis Generation in the Computational Model for Visual Word Recognition", IEEE Expert, Fall 1986, pp. 63-70.
- J. Mariani "Recent Advances in Speech Processing", ICASSP '89 23-26 May 1989 Scottish Exhibition, pp. 429-440.
- L. R. Rabiner "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition", Proceedings of the IEEE, vol. 77, No. 2, Feb. 1989, pp. 257-286.
- S. N. Srihari and G. W. Zack "Document Image Analysis", Proc. 8th. Intl. Conf. Pattern Recognition, Paris, Oct. 1986, pp. 434-436.
- A. Kundu, Y. He and P. Bahl "Recognition of Handwritten Word: First and Second Order Hidden Markov Model Based Approach" Pattern Recognition, vol. 22, No. 3, 1989, pp. 283-297.
- T. Pavlidis "A Vectorizer and Feature Extractor for Document Recognition", Computer Vision, Graphics and Image Process, vol. 35, 1986, pp. 111-127.
- L. R. Rabiner and S. E. Levinson "A Speaker-Independent, Syntax-Directed, Connected Word Recognition System Based on Hidden Markov Models and Level Building", IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-33, No. 3, Jun. 1985, pp. 561-573.
- A. K. Jain and R. C. Dubes "Algorithms for Clustering Data", Chapter 3, Prentice Hall 1988, pp. 55-142.
- R. O. Duda and Peter E. Hart "Pattern Classification and Scene Analysis", Wiley 1973, pp. 10-43.
- A. G. Konheim "Cryptography -A Primer", Wiley 1981, pp. 14-27.
- S. E. Levinson, L. R. Rabiner and M. M. Sondhi "An Introduction to the Application of the Theory of Probabilistic Functions of a Markov Process to Automatic Speech Recognition", Bell System Technical Journal, vol. 62, No. 4, Apr. 1983, pp. 1035-1074.
- H. S. Baird "Document Image Defect Models", Proc. IAPR Workshop on SSPR, Murray Hill, NJ 13-15 Jun. 1990, 1-9.
- D-S. Lee et al. "A Structural Approach to Recognize Hand-printed and Degraded Machine-printed Characters", IAPR Workshop on Syntactic and Structural Pattern Recognition, Jun. 1990, 256-272.
- J. J. Hull and S. N. Srihari "A Computational Approach to Visual Word Recognition: Hypothesis Generation and Testing", Proc. IEEE Conf. Comput. Vision Pattern Recognition, 1986, 156-161.
- H. Baird "Global-to-local layout analysis", Proc. IAPR Workshop on Syntactic and Structural Pattern Recog., France, Sep. 1988.