Identification and Segmentation of Finely Textured and Solid Regions of Binary Images

A portion of the disclosure of this patent document contains material which 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 Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever.

The invention relates generally to image processing, and more specifically to a technique for discriminating between regions that are finely textured or solid and other regions such as those containing text or line graphics.

There are several applications where it is important to determine quickly whether an image contains regions of fine texture or regions that are fairly solidly dark. Such regions include halftones, stipples, and dark areas with light lettering.

For example, problems can arise with printers whose output is at a different resolution from the input scanner, since techniques for converting the resolution are sensitive to texture. A technique that works well on text will generally do poorly on halftones or stipples, and vice versa. If the halftone regions are identified, and if the frequency and screen angle are known, then appropriate techniques can be selected for different portions of the image, and the halftone regions can be resolution converted with an acceptably low level of aliasing artifacts.

Problems also arise for scanners. When scanning a halftoned binary image, beating can occur between the repeat frequency in the image and the size of an integral number of scanned pixels. The result is aliasing, in which a low frequency Moire pattern will be observed in the scanned image. In order to prevent this, it is desirable to use a gray scale scanner, and remove halftoning prior to thresholding. Techniques for removing halftone patterns work best if the halftone frequency and screen angle are known.

It is sometimes necessary for gross segmentation to occur prior to using some segmentation software. Some aspects of segmentation can be accomplished by building a connected component representation of the binary image, such as a line adjacency graph (LAG), and then processing that data structure. However, if one tried to build a LAG from a finely textured screen of appreciable size (say 2 inches square or larger), the storage requirements and computational time might well be excessive. Similarly, it is necessary for segmentation to occur prior to using recognition software. If a halftoned region of a binary image were sent to OCR or graphics vectorization software, it could break the program.

Depending on the threshold setting and resolution of the input scanners, and on the quality of the output printer and number of copy generations, a printed version of the binary image of a regular textured region may show little or none of the details of the texturing of the original binary image. Because of scanning and printing operations, the contrast at the 1-5 line pairs per mm size is often greatly increased, with dark textures becoming solid black, and light textures becoming much lighter or even white (perhaps with some random dot noise). Thus, in many situations, it cannot be assumed that evidence of the original texturing will remain in the binary image under analysis.

The present invention provides a robust and computationally efficient technique for identifying and separating regions that either have fine texture or are fairly solidly dark in a binary image. The invention properly separates halftones without requiring any knowledge of the frequency and screen angle, and works well on improperly thresholded halftones (too light or too dark).

In brief, the invention contemplates the use of a series of transformations, such as thresholded reduction operations and morphological operations to transform regions according to their texture.

Threshold reductions and morphological operations will be defined and discussed in detail below. A thresholded reduction entails mapping a rectangular array of pixels onto a single pixel, whose value depends on the number of ON pixels in the rectangular array and a threshold level. Morphological operations use a pixel pattern called a structuring element (SE) to erode, dilate open, or close an image.

A method according to one aspect of the invention comprises converting the finely textured and solid regions to solid ON (black) pixels, clearing the text and graphics regions, optionally filling the holes (if any) in the solid areas to provide a mask, and then, if necessary, adjusting the scale of the mask to conform to the original image or extracting relevant mask coordinates. The halftone separation is obtained by ANDing the mask and the original image while the remainder is obtained by XORing the halftone separation and the original image.

The image is subjected to a first set of operations that eliminates OFF pixels that are near ON pixels, which tends to thicken text and lines and solidify textured regions Solid regions remain solid. The first set of operations preferably includes a thresholded reduction with general darkening (threshold LEVEL=1), followed by a close operation.

The image is then subjected to a second set of operations that eliminates ON pixels that are near OFF pixels. This thins out and eliminates the previously thickened text and lines, but leaves the previously solidified textured regions substantially intact. The second set of operations preferably includes a thresholded reduction with general lightening (threshold LEVEL=4) followed by an open operation.

Filling the holes preferably includes an iterated series of erosions and dilations to fill 8-connected regions to the minimum enclosing rectangular regions. The method may be controlled by counting the number of filled regions, and performing a series of open operations to reduce the number of filled regions.

The present invention is computationally simple since the operations are essentially local. The invention is also fast. This is in part because the operations entail few computations per pixel, and in part because many of the operations are carried out on reduced images, which means there are fewer pixels to process.

A further understanding of the nature and advantages of the present invention may be realized by reference to the remaining portions of the specification and the drawings.

FIG. 1A is a block diagram of an image scanning and processing system incorporating the present invention;

FIG. 1B is a high level flow diagram showing the method of separating finely textured and solid regions in an image;

FIG. 1C shows a representative binary image;

FIG. 2A is a flow diagram illustrating the filling of finely textured regions of the image;

FIGS. 2B-D show the results at different stages of the filling operation;

FIG. 3A is a flow diagram illustrating the removal of text and line graphics from the image;

FIGS. 3B-D show the results at different stages of the removal operation;

FIG. 4 is a flow diagram illustrating the filling of 8-connected regions in the mask;

FIG. 5 is a flow diagram of an optional procedure for testing and removing small rectangular areas in the mask;

FIG. 6A is a flow diagram illustrating the adjustment of the mask size to conform to the size of the original textured area;

FIG. 6B shows the full-size mask;

FIG. 7A is a flow diagram illustrating the use of the mask to obtain the separation of the textured and non-textured regions;

FIGS. 7B-D show the halftone separation, the text separation, and the mask coverage;

FIGS. 8A-B are flow diagrams illustrating alternatives to the thresholded reductions used to fill the finely textured regions and remove text and line graphics;

FIG. 9 is a flow diagram illustrating an alternative to expanding the mask;

FIG. 10 is a flow diagram illustrating the filling of 4-connected regions in the mask; and

FIG. 11 is a block diagram of special purpose hardware for performing image reductions and expansions.

The present discussion deals with 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 for processing will sometimes be referred to as the original image.

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. While the techniques of the present invention could be applied to negative images as well, the discussion will be in terms of black on white.

A "solid region" of an image refers to a region extending many pixels in both dimensions within which substantially all the pixels are ON.

A "textured region" of an image refers to a region that contains a relatively fine-grained pattern. Examples of textured regions are halftoned or stippled regions.

"Text" refers to portions of a document or image containing letters, numbers, or other symbols including non-alphabetic linguistic characters.

"Line graphics" refers to portions of a document or image composed of graphs, figures, or drawings other than text, generally composed of horizontal, vertical, and skewed lines having substantial run length as compared to text. Graphics could range from horizontal and vertical lines in an organization chart to more complicated horizontal, vertical, and skewed lines in engineering drawings.

A "mask" refers to an image, normally derived from an original 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 don't correspond to regions of interest.

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

"Reduction" is a scale operation characterized by a SCALE factor N and a threshold LEVEL M. Reduction with SCALE=N entails dividing the source image into NxN 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.

"Subsampling" is an operation wherein the source image is subdivided into smaller (typically square) elements, and each element in the source image is mapped to a smaller element in the destination image. The pixel values for each destination image element are defined by a selected subset of the pixels in the source image element. Typically, subsampling entails mapping to single pixels, with the destination pixel value being the same as a selected pixel from the source image element. The selection may be predetermined (e.g. upper left pixel) or random.

A "4-connected region" is a set of ON pixels wherein each pixel in the set is laterally or vertically adjacent to at least one other pixel in the set.

An "8-connected region" is a set of ON pixels wherein each pixel in the set is laterally, vertically, or diagonally adjacent to at least one other pixel in the set.

A number of morphological operations map a source image onto an equally sized destination image according to a rule defined by a pixel pattern called a structuring element (SE). The SE is defined by a center location and a number of pixel locations, each having a defined value (ON or OFF). Other pixel positions, referred to as "don't care", are ignored. 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 "hit-miss" SE refers to an SE that specifies at least one ON pixel and at least one OFF pixel.

"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 ON and OFF pixels in the SE and the underlying pixels in the source image.

"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 SE's used for dilation typically have no OFF pixels.

"Opening" is a morphological operation that consists of an erosion followed by a dilation. The result is to replicate the SE in the destination image for each match in the source image.

"Closing" is a morphological operation consisting of a dilation followed by an erosion

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.

FIG. 1A is a block diagram of an image analysis system 1 within which the present invention may be embodied. The basic operation of system 1 is to extract or eliminate certain characteristic portions of a document 2. To this end, the system includes a scanner 3 which digitizes the document on a pixel basis, and provides a resultant data structure, typically referred to as an image. Depending on the application, the scanner may provide a binary image (a single bit per pixel) or a gray scale image (a plurality of bits per pixel). The image contains the raw content of the document, to the precision of the resolution of the scanner. The image may be sent to a memory 4 or stored as a file in a file storage unit 5, which may be a disk or other mass storage device.

A processor 6 controls the data flow and performs the image processing. Processor 6 may be a general purpose computer, a special purpose computer optimized for image processing operations, or a combination of a general purpose computer and auxiliary special purpose hardware. If a file storage unit is used, the image is transferred to memory 4 prior to processing. Memory 4 may also be used to store intermediate data structures and possibly a final processed data structure.

The result of the image processing, of which the present invention forms a part, can be a derived image, numerical data (such as coordinates of salient features of the image) or a combination. This information may be communicated to application specific hardware 8, which may be a printer or display, or may be written back to file storage unit 5.

FIG. 1B is a high level flow diagram showing the major steps for creating a mask that covers textured and solid regions in an original binary image. The original binary image may include regions of fine texturing such as stippling or halftoning, solid black regions, and regions that are neither solid nor finely textured, such as those containing text and line graphics. As will be seen below, the image processing that occurs tends to leave originally solid regions of sufficient extent alone and the solid regions will appear as part of the mask. Thus the discussion below will sometimes be silent as to the solid regions. Non-textured regions that are not solid will sometimes be referred to as other regions.

FIG. 1C shows a representative binary image scanned at 300 dots/inch having text and halftone regions. Certain other views of the images at intermediate stages of processing are magnified 2.times. relative to FIG. 1B. References to magnification in the following discussion are relative to FIG. 1B and are not to be confused with the expansion and reduction scale operations.

In brief, the original binary image is subjected to a series of operations that may include logical, scale, and morphological operations. The image is first processed so that the textured regions are converted to solid black, with the possibility of some holes (step 10). Text and graphics regions are then cleared (step 12). The resulting image may then be subjected to a number of optional steps wherein any holes in the solid areas are filled (step 15). The process can be controlled (optionally) to determine if there are too many solid regions, and if so, to remove the small regions (step 17).

The solid regions are then used to characterize the halftone regions. One option is to expand the solid regions back to their original size, thus forming a mask for separation of finely textured regions from the rest of the image (step 18). The mask and the original binary image can then be combined to provide the separations of the textured and non-textured regions (step 20). Alternatively, the corner positions of the unexpanded solid regions can be scaled and used to determine the textured regions in the original binary image.

FIG. 2A is an expanded flow diagram illustrating the steps within step 10 (blackening textured regions). The original image is subjected to a set of operations that eliminate OFF pixels that are near ON pixels. While text and lines are thickened, they tend to retain their general character. However, as the small dots in the textured regions expand, they coalesce to form large masses and thereby solidify the formerly textured area. The solid regions remain solid. As will be seen below, subsequent processing can reverse thickening of the characters and lines, but not the solidification of the now solid regions.

The original binary image is twice reduced with SCALE=2 and LEVEL=1 (steps 22 and 23). Each reduction step entails dividing the image (original binary image for step 22 and reduced image for step 23) into 2.times.2 squares, each of which maps onto a single pixel in the destination image. The destination pixel is turned ON if one or more of the pixels in the square is ON, and turned OFF if no pixels in the square are ON. The result is now an image reduced by a linear factor of 4 and having the textured regions darkened.

This reduced image is then subjected to a close operation (step 25) to finish the solidification of the textured region. The close operation consists of a dilation and an erosion, preferably with a solid 2.times.2 SE (i.e., an SE that is a 2.times.2 square of ON pixels). The result of the close operation is invariant as to which of the pixels in the SE is designated the center.

FIGS. 2B and 2C show the results (at 2.times. magnification) of the two reductions with LEVEL=1. FIG. 2D shows the result (at 2.times. magnification) of the close operation. These two reductions and the close operation cause the textured and solid regions to fill. They also cause features in the other regions to fill somewhat However, the spatial extent of regions of ON pixels in the other regions is much smaller than the filled textured regions.

FIG. 3A is an expanded flow diagram illustrating the steps within step 12 (removing text and graphics). The image resulting from step 10 is subjected to a set of operations that eliminate ON pixels that are near OFF pixels. This thins and eliminates the previously thickened features in the other regions (text and lines), but does not materially affect the now solid regions.

The removal of text and graphics is accomplished by successive reductions (steps 32 and 33), again each with SCALE=2, but with LEVEL=4 (all four pixels must be ON to turn on the destination pixel). The resulting image, now reduced by a linear factor of 16, contains only a few isolated ON pixels within the other regions. This image is subjected to an open operation (step 35), preferably with the same solid 2.times.2 SE as in close operation 25.

FIGS. 3B and 3C show the results (at 2.times. magnification) of the two reductions with LEVEL=4. FIG. 3D shows the result (at 2.times. magnification) of the OPEN operation. The net result is to eliminate the text while leaving the filled textured region intact.

FIG. 4 is an expanded flow diagram illustrating the steps within step 15 (filling holes in solid areas). A presently preferred technique for filling holes within the remaining dark regions is to fill all 8-connected regions to the smallest possible enclosing rectangle. This is accomplished by an iterated sequence of erosions and dilations using two diagonal SE's 42 and 43. SE 42 has two ON pixels, one to the immediate right of the center and one immediately beneath the center. SE 43 has two ON pixels, one at the center and one diagonally down to the right.

The input image (reduced by a scale factor of 16 as described above) is copied (step 45), with one copy reserved for later use and one copy being a work copy subject to succeeding operations. The work copy is first eroded (step 47) with SE 42, and then dilated (step 50) with SE 43. The result of this erosion and dilation is subjected to a logical OR (step 55) with the copy reserved at copy step 45. The result of the logical OR is copied (step 57), with one copy being reserved for use and the other being a work copy. The work copy is eroded (step 60) with SE 42, dilated (step 62) with SE 43. The resulting image and the copy reserved at copy step 57 are subjected to a logical OR (step 65). The resultant iterated image and the copy of the input image reserved at step 45 are subjected to an exclusive OR. If the iterated image has not changed (the XOR of the two images contains no ON pixels), the process is complete. If the iterated image has changed (the XOR of the two images contains at least one ON pixel), the iterated image is communicated back and subjected to steps 45 through 67. The cycle repeats until the iterated image agrees with the last version reserved at copy step 45. In certain applications, it may be desired to perform a small number of iterations to effect only a partial fill.

FIG. 5 is an expanded flow diagram illustrating the steps within step 17 (removing excess solid regions). As noted above, step 17 is optional. In most cases it is not needed. There are, however, circumstances where very large characters or very dense line graphics will not be entirely removed by the preceding steps. Step 17 will remove such exceptional items and flag their existence for further processing.

The basic procedure entails determining whether the number of solid regions is greater than some maximum level Max1, and if so, attempting to remove the smaller of the regions. The number of solid rectangles is determined by looking for upper left corners. This is done by erosion with an SE 82 that consists of a 2.times.2 array of pixels with the center pixel being ON in the lower right corner and the three remaining pixels being OFF. Removal of excessive solid regions is carried out by open operations with increasingly large SE's consisting of square arrays with all pixels ON.

Upon entry into this procedure, the SE with solid ON pixels for possible removal of solid regions is initialized (step 85) at a 2.times.2 size. The input image is copied (step 87), with one copy reserved for possible output and the other transmitted for determination the solid areas. The number of solid areas is determined by an erosion (step 90) using SE 82. The effect of the erosion step is to produce a single ON pixel in the destination image at the location of the upper left corner of each solid rectangle in the source image. The ON pixels are counted (step 92), and the number compared to Max1 (step 95). In the event that the number of ON pixels, which corresponds to the number of solid regions in the original image is less than or equal to Max1, the original image reserved at copy step 87 is output.

Otherwise, further removal is effected as follows. The size of the SE is increased by 1 on a side (step 97), and the new SE size is compared to a predetermined maximum Max2 (step 100). If the size of the SE exceeds Max2, the image is output with a warning. Otherwise, the image is subjected to an OPEN operation (step 102) with the newly increased SE. The result of this is iterated until the number of solid regions is below Max1 or the SE has reached Max2.

FIG. 6A is an expanded flow diagram illustrating the steps within step 18 (expanding the mask to original size). The result of the previous operations is a mask at reduced scale. It is possible to expand the solid regions back to the original size to form the mask for separation of the finely textured (and solid) regions from the rest of the image. The result of step 12 (two reductions and an open) was to relatively reduce the size of the filled regions. This is compensated by first expanding the image with SCALE=2 (step 112), and then dilating the image (step 115) using a 3.times.3 SE 117 with the all pixels ON and the SE center in the middle. Dilating with SE 117 causes the boundary of each rectangle of ON pixels to be expanded outwardly by one pixel. The resulting image is then expanded with SCALE=8 (step 117), returning to full scale. FIG. 6B shows the mask at full scale.

FIG. 7A is an expanded flow diagram illustrating the steps within step 20 (image separation). With the mask thus produced, segmentation is easily accomplished. The original binary image is copied (step 130) and ANDed with the mask (step 135) to produce the "halftone separation" (i.e. the textured and solid parts of the image). This is copied (step 137) and the result is subjected to an exclusive OR (step 140) with a copy of the original binary image to produce the other parts of the image (text and line graphics). It is noted that the copying steps are not necessary if the result of the logical operation at any given stage is stored in a new array in memory rather than overwriting the operand.

FIG. 7B shows the halftone separation achieved by ANDing the mask and the original image and FIG. 7C shows the text separation achieved by XORing the halftone separation and the original image. FIG. 7D shows the result of XORing the mask and the halftone separation, which demonstrates the mask coverage.

FIG. 8A is an expanded flow diagram illustrating an alternative to the use of thresholded reductions with LEVEL=1 to fill the textured areas. More particularly, the image is dilated with a solid 2.times.2 SE (all pixels ON) (step 150) and the resultant image subsampled by choosing one pixel in each 2.times.2 square to form a reduced image (step 152). The sub-sampling may be accomplished on a row basis by discarding every other line, and on the column basis by use of a lookup table in a manner similar to that described below in connection with performing fast thresholded reductions.

In principle one could also use a series of close operations in an attempt to solidify a finely textured region. However, the use of one or more thresholded reductions (or dilation and subsampling) has at least two advantages. First, because the texture scale is not known a priori, it cannot be determined how large an SE to use in the close operation. An SE that is too small to bridge adjacent parts in the textured region would not change the image, and the close operation would fail. Thus, while the use of the close would be locally all or nothing, the use of a reduction with LEVEL=1 (or a dilation and subsampling) invariably results in a darkening of the texture. Second, the use of reductions before close allows the close operation to be carried out at a reduced scale. Comparable operations at full scale are much slower computationally than those at a reduced scale (roughly by the third power of the linear scale factor). Therefore, all subsequent operations at reduced scale are much faster.

FIG. 8B is an expanded flow diagram illustrating an alternative to the use of thresholded reductions with LEVEL=4 to remove text and line graphics. The image is first eroded with a solid 2.times.2 SE (step 155) and the resultant image sub-sampled by choosing one pixel in each 2.times.2 square to form a reduced image (step 157).

The above comments in connection with filling the textured regions apply to removing text and line features. In principle one could also use a series of open operations in an attempt to remove such features. However, the use of one or more thresholded reductions (or erosion and subsampling) has at least two advantages. First, because the feature scale is not known a priori, it cannot be determined how large an SE to use in the open operation. An SE that is too small to eliminate features in the textured region would not change the image, and the open operation would fail. Thus, while the use of the open would be locally all or nothing, the use of a reduction with LEVEL=4 (or an erosion and subsampling) invariably results in a removal or thinning of the feature. Second, the use of reductions before open allows the open operation to be carried out at a reduced scale.

The selection of a particular size for the SE is empirically based, taking into consideration such factors as scanner resolution. However, the decision tends to be rather straightforward, entailing a minimum amount of experimentation.

Although the thresholded reductions are considered advantageous for the filling and removal steps, the same result is likely to be achieved by doing a dilation with a solid 8.times.8 SE and an erosion with a solid 32.times.32 SE. However, the computation is likely to take much longer.

FIG. 9 is a flow diagram illustrating an alternative to expansion of the solid portions to form the mask. Instead of expanding the solid regions to form the mask, it is possible to extract the coordinates of the solid regions (step 160) and scale the coordinates to full size (step 161). This provides a compact representation that allows convenient storage of the mask information.

The locations of the corners of each solid rectangular portion can be extracted by eroding the copies of the mask with a series of four SE's 162(ULC), 162(URC), 162(LLC), and 162(LRC). SE 162(ULC) is the same as SE 82, and picks out the upper left corner when it is used to erode a rectangle. The other SE's pick out the other corners.

For rectangular regions, it is useful, but not necessary, to extract the corner coordinates in a known order; the coordinates alone dictate the association of corners with rectangular regions. However, for non-rectangular regions, where there are more than four corners for each connected region of the mask (and there may be holes), the corners must be maintained in the order they are encountered by tracing around the periphery. Although possible, this may be sufficiently complicated that it is preferable to use the mask itself.

FIG. 10 is an expanded flow diagram illustrating an alternative technique for filling holes in solid areas. The technique is to fill all 4-connected regions (rather than 8-connected regions as described above) to the smallest enclosing rectangle. This is accomplished by an iterated sequence of erosions and logical OR's using 4 SE's 171, 172, 173, and 174. Each of the SE's contains three ON pixels at three of the four corners of a 2.times.2 square with the remaining corner being a "don't care" position which also defines the center position of the SE. The input image (reduced by a scale factor of 16 as described above) is copied (step 175), with one copy reserved for later use and one copy being a work copy. The work copy is eroded (step 176) with SE 171, and then ORed (step 177) with the work copy. The resulting image is then copied (step 178) with one copy being reserved for use and the other being a work copy. The work copy is eroded (step 179) with SE 172, and the result is ORed (step 180) with the work copy. The sequence is repeated with copy step 181, erosion step 182 with SE 173 and logical OR step 183, followed by copy step 184, erosion step 185 with SE 174, and logical OR step 186. The result is an iterated image which is subjected to an exclusive OR (step 190) with the image input to copy step 175. If the iterated image is not changed (the XOR of the two images contains no ON pixels), the process is complete If the iterated image has changed (the XOR of the two images contains at least one ON pixel), the iterated image is communicated back and subjected to steps 175 through 186. The complete cycle repeats until the iterated image agrees with the last version reserved at copy step 175. In certain applications, it may be desired to perform a small number of iterations to effect only a partial fill.

One requirement of efficient segmentation, is that thresholded reduction must be done quickly. Suppose it is desired to reduce an image by a factor of two in the vertical direction. One way to do this is to use a raster operation (bitblt--bit block transfer) to logically combine the odd and even rows, creating a single row of the reduced image for each pair of rows in the original. The same procedure can then be applied to the columns of the vertically squashed image, giving an image reduced by a factor of two in both directions.

The result, however, depends on the logical operations of the horizontal and vertical raster operations. Obtaining a result with LEVEL=1 or 4 is straightforward. If an OR is used for both raster operation orientations, the result is an ON pixel if any of the four pixels within the corresponding 2.times.2 square of the original were ON. This is simply a reduction with LEVEL=1. Likewise, if an AND for both raster operation orientators, the result is a reduction with LEVEL=4, where all four pixels must be ON.

A somewhat different approach is used to obtain a reduction with LEVEL=2 or 3. Let the result of doing a horizontal OR followed by a vertical AND be a reduced image R1, and let the result from doing a horizontal AND followed by a vertical OR be a reduced image R2. A reduction with LEVEL=2 is obtained by ORing R1 with R2, and a reduction with LEVEL=3 is obtained by ANDing R1 with R2.

The procedure may not be computationally efficient if implemented as described above. On some computers, such as Sun workstations, raster operations are done in software. The image is stored as a block of sequential data, starting with the first row of the image, moving left-to-right, then the second row, etc. Consequently, the raster operations between rows are fast, because 16 or 32 bits in two words can be combined in one operation. But to perform a raster operation between two columns, the corresponding bits must be found, two bits at a time (one from each column), before the logical operations can be done. It turns out that the time, per pixel, to do the vertical raster operations is at least 25 times greater than the horizontal ones. In fact, when the algorithm is implemented entirely with raster operations, over 90 percent of the time is devoted to the vertical operations.

Fortunately, there is a simple and very fast way to implement the logical operations between columns. Rather than use column raster operations, take 16 sequential bits, corresponding to 16 columns in one row. These 16 bits can be accessed as a short integer. These 16 bits are used as an index into a 2.sup.16 -entry array (i.e. a lookup table) of 8-bit objects. The 8-bit contents of the array give the result of ORing the first bit of the index with the second, the third bit with the fourth . . . and on to the 15th bit with the 16th. Actually, two arrays are needed, one for ORing the 8 sets of adjacent columns, and one for ANDing the columns. It should be understood that the numerical example is just that, an example. It is also possible to implement this as a 2.sup.8 -entry array of 4-bit objects, or any one of a number of other ways.

The use of lookup tables to implement column logical operations is about as fast, per pixel, as Sun's row raster operations. A 1000.times.1000 pixel image can be reduced on a Sun 3/260, with either LEVEL=1 or 4, to a 500.times.500 pixel image in 0.1 seconds. On a Sun 4/330, the operation takes about 0.04 second.

As discussed above, 2.times.2 reductions require a first logical operation between rows followed by a second, possibly different, logical operation between columns. Moreover, some threshold levels require two intermediate reduced images which are then combined. The table lookup technique for column operations can become cumbersome if it is desired to have a very wide pixelword. Either the table becomes enormous or one needs special techniques of looking up parts of the wide pixelword in multiple parallel tables. The latter, while clearly superior, does require some way to use portions of the data word as memory addresses, which may not otherwise be necessary.

FIG. 10 is a logic schematic of specialized hardware for performing a logical operation between vertically adjacent 2Q-bit pixelwords and a pairwise bit reduction of the resulting 2Q-bit pixelword (bits 0 through 2Q-1). Although the drawing shows a 16-pixel word, the benefits of this hardware would become manifest for much longer pixelwords where the lookup table technique has become cumbersome. A 512-bit pixelword is contemplated, since a line of image would represent only a few pixelwords.

The reduction of the two pixelwords occurs in two stages, designated 200 and 202. In the first stage, a vertically adjacent pair of pixelwords is read from a first memory 203, and the desired first logical operation is carried out between them. The desired second logical operation is then carried out between the resulting pixelword and a version of the pixelword that is shifted by one bit. This provides a processed pixelword having the bits of interest (valid bits) in every other bit position. In the second stage, the valid bits in the processed pixelword are extracted and compressed, and the result stored in a second memory 204. Memory 203 is preferably organized with a word size corresponding to the pixelword size. Memory 204 may be organized the same way.

The preferred implementation for stage 200 is an array of bit-slice processors, such as the IDT 49C402 processor, available from Integrated Device Technology. This specific processor is a 16-bit wide device, each containing 64 shiftable registers. Thirty-two such devices would be suitable for a 512-bit pixelword. For simplification, a 16-bit system with four registers 205, 206, 207, and 208 is shown. Among the processor's operations are those that logically combine the contents of first and second registers, and store the result in the first. The processor has a data port 215, which is coupled to a data bus 217.

Second stage 202 includes first and second latched transceivers 220 and 222, each half as wide as the pixelword. Each transceiver has two ports, designated 220a and 220b for transceiver 220 and 222a and 222b for transceiver 222. Each transceiver is half as wide as the pixelword. Ports 220a and 222a are each coupled to the odd bits of data bus 217, which correspond to the bits of interest. Port 220b is coupled to bits 0 through (Q-1) of the data bus, while port 222b is coupled to bits Q through (2Q-1). The bus lines are pulled up by resistors 125 so that undriven lines are pulled high.

Consider the case of a 2.times.2 reduction with LEVEL=2. The sequence of operations requires that (a) a vertically adjacent pair of pixelwords be ANDed to form a single 2Q-bit pixelword, adjacent pairs of bits be ORed to form a Q-bit pixelword, and the result be stored; (b) the vertically adjacent pair of pixelwords be ORed, adjacent bits of the resultant 2Q-bit pixelword be ANDed, and the resultant Q-bit pixelword be stored; and (c) the two Q-bit pixelwords be ORed.

To effect this, a pair of vertically adjacent pixelwords are read from first memory 203 onto data bus 217 and into registers 205 and 206. Registers 205 and 206 are ANDed and the result stored in registers 207 and 208. The content of register 208 is shifted one bit to the right, registers 207 and 208 are ORed, and the result is stored in register 208. Registers 205 and 206 are ORed, and the result stored in registers 206 and 207. The content of register 207 is right shifted by one bit, registers 206 and 207 are ANDed, and the result stored in register 207.

At this point, register 207 contains the result of ORing the two pixelwords and ANDing pairs of adjacent bits, while register 208 contains the result of ANDing the pixelwords and ORing pairs of adjacent bits. However, registers 207 and 208 contain the valid bits in the odd bit positions 1, 3, . . . (2Q-1). For a reduction with LEVEL=2, registers 207 and 208 are ORed and the result is made available at processor data port 215 which is coupled to data bus 217.

The odd bits of the data bus are latched into transceiver 220 through port 220a, resulting in a Q-bit pixelword with the valid bits in adjacent positions. Although this Q-bit entity could be read back onto the bus and transferred to memory 204, it is preferable to use both latches. Thus, two new pixelwords (horizontally adjacent to the first two) are processed at stage 200 as described above, the result is made available at processor data port 215, and is latched into transceiver 222 through port 222a. The contents of the two transceivers are then read out through ports 220b and 222b onto data bus 217 in order to provide a 2Q-bit pixelword that represents the reduction of four 2Q-bit pixelwords. The result is transferred to second memory 204. This overall sequence continues until all the pixelwords in the pair of rows has been processed. Once the pair of rows has been processed, subsequent pairs are similarly processed.

As mentioned above each bit-slice processor has 64 registers. Accordingly, since memory accesses are more efficient in a block mode, faster operation is likely to result if 8 pairs of pixelwords are read from memory 203 in a block, processed as discussed above, stored in the processor's registers, and written to memory 204 in a block.

Image enlargement is similar, but the steps are executed in the reverse order. First, the processor reads a pixelword and sends the left half through port 220b of transceiver 220. This is read onto the bus through port 220a. Only every other pixel in the resulting word on the bus will initially be valid, so the processor will need to validate all the pixels using a sequence of shifts and logic operations. Since resistors 225 pull up all the bus lines that are not driven, each undriven line, all the even bits in this case, will be 1's. This expanded pixelword, which alternates 1's with valid data, is read into two registers, the content of one register is shifted one place, and the registers are logically ANDed. Everywhere there was a 0 in an odd bit, there will be 00 in an even/odd pair. None of the other bits will be affected This pixelword is then written to two vertically adjacent words in the expanded image. This process is repeated for the right half of the pixelword using the transceiver 222. The processor expands the entire row one pixelword at a time and the entire image one row at a time.

A current embodiment of the invention is implemented in software on a digital computer. Appendix 1 (.COPYRGT. 1988, Unpublished Work, Xerox Corporation) provides a source code program for implementation of this embodiment. The program is in the "C" language, well known to those of skill in the art. The program has been demonstrated on a Sun Workstation, although it will be apparent to those of skill in the art that a wide variety of programming languages and hardware configurations could readily be used based on this disclosure without departing from the scope of the invention.

In conclusion it can be seen that the invention provides efficient and effective separation of textured regions from non-textured regions in an image.

While the above is a complete description of the preferred embodiment of the invention, various modifications are possible. The above discussion of alternatives to certain steps is representative of the flexibility of the invention. Furthermore, all the above operations could be done on an image that is first reduced. Thus, much of the image computation would occur at a further reduced scale. Moreover, while the specific original image illustrated above was derived from a document scanned at 300 pixels/inch, the invention applies to documents scanned at any other resolution.

Therefore the above description and illustrations should not be taken as limiting the scope of the invention, which is defined by the claims. ##SPC1##