- COPYRIGHT AND LICENSE
jbclass.c These are functions for unsupervised classification of collections of connected components -- either characters or words -- in binary images. They can be used as image processing steps in jbig2 compression. Initialization JBCLASSER *jbRankHausInit() [rank hausdorff encoder] JBCLASSER *jbCorrelationInit() [correlation encoder] JBCLASSER *jbCorrelationInitWithoutComponents() [ditto] static JBCLASSER *jbCorrelationInitInternal() Classify the pages l_int32 jbAddPages() l_int32 jbAddPage() l_int32 jbAddPageComponents() Rank hausdorff classifier l_int32 jbClassifyRankHaus() l_int32 pixHaustest() l_int32 pixRankHaustest() Binary correlation classifier l_int32 jbClassifyCorrelation() Determine the image components we start with l_int32 jbGetComponents() l_int32 pixWordMaskByDilation() l_int32 pixWordBoxesByDilation() Build grayscale composites (templates) PIXA *jbAccumulateComposites PIXA *jbTemplatesFromComposites Utility functions for Classer JBCLASSER *jbClasserCreate() void jbClasserDestroy() Utility functions for Data JBDATA *jbDataSave() void jbDataDestroy() l_int32 jbDataWrite() JBDATA *jbDataRead() PIXA *jbDataRender() l_int32 jbGetULCorners() l_int32 jbGetLLCorners() Static helpers static JBFINDCTX *findSimilarSizedTemplatesInit() static l_int32 findSimilarSizedTemplatesNext() static void findSimilarSizedTemplatesDestroy() static l_int32 finalPositioningForAlignment() Note: this is NOT an implementation of the JPEG jbig2 proposed standard encoder, the specifications for which can be found at http://www.jpeg.org/jbigpt2.html. (See below for a full implementation.) It is an implementation of the lower-level part of an encoder that: (1) identifies connected components that are going to be used (2) puts them in similarity classes (this is an unsupervised classifier), and (3) stores the result in a simple file format (2 files, one for templates and one for page/coordinate/template-index quartets). An actual implementation of the official jbig2 encoder could start with parts (1) and (2), and would then compress the quartets according to the standards requirements (e.g., Huffman or arithmetic coding of coordinate differences and image templates). The low-level part of the encoder provided here has the following useful features: - It is accurate in the identification of templates and classes because it uses a windowed hausdorff distance metric. - It is accurate in the placement of the connected components, doing a two step process of first aligning the the centroids of the template with those of each instance, and then making a further correction of up to +- 1 pixel in each direction to best align the templates. - It is fast because it uses a morphologically based matching algorithm to implement the hausdorff criterion, and it selects the patterns that are possible matches based on their size. We provide two different matching functions, one using Hausdorff distance and one using a simple image correlation. The Hausdorff method sometimes produces better results for the same number of classes, because it gives a relatively small effective weight to foreground pixels near the boundary, and a relatively large weight to foreground pixels that are not near the boundary. By effectively ignoring these boundary pixels, Hausdorff weighting corresponds better to the expected probabilities of the pixel values in a scanned image, where the variations in instances of the same printed character are much more likely to be in pixels near the boundary. By contrast, the correlation method gives equal weight to all foreground pixels. For best results, use the correlation method. Correlation takes the number of fg pixels in the AND of instance and template, divided by the product of the number of fg pixels in instance and template. It compares this with a threshold that, in general, depends on the fractional coverage of the template. For heavy text, the threshold is raised above that for light text, By using both these parameters (basic threshold and adjustment factor for text weight), one has more flexibility and can arrive at the fewest substitution errors, although this comes at the price of more templates. The strict Hausdorff scoring is not a rank weighting, because a single pixel beyond the given distance will cause a match failure. A rank Hausdorff is more robust to non-boundary noise, but it is also more susceptible to confusing components that should be in different classes. For implementing a jbig2 application for visually lossless binary image compression, you have two choices: (1) use a 3x3 structuring element (size = 3) and a strict Hausdorff comparison (rank = 1.0 in the rank Hausdorff function). This will result in a minimal number of classes, but confusion of small characters, such as italic and non-italic lower-case 'o', can still occur. (2) use the correlation method with a threshold of 0.85 and a weighting factor of about 0.7. This will result in a larger number of classes, but should not be confused either by similar small characters or by extremely thick sans serif characters, such as in prog/cootoots.png. As mentioned above, if visual substitution errors must be avoided, you should use the correlation method. We provide executables that show how to do the encoding: prog/jbrankhaus.c prog/jbcorrelation.c The basic flow for correlation classification goes as follows, where specific choices have been made for parameters (Hausdorff is the same except for initialization): // Initialize and save data in the classer JBCLASSER *classer = jbCorrelationInit(JB_CONN_COMPS, 0, 0, 0.8, 0.7); SARRAY *safiles = getSortedPathnamesInDirectory(directory, NULL, 0, 0); jbAddPages(classer, safiles); // Save the data in a data structure for serialization, // and write it into two files. JBDATA *data = jbDataSave(classer); jbDataWrite(rootname, data); // Reconstruct (render) the pages from the encoded data. PIXA *pixa = jbDataRender(data, FALSE); Adam Langley has built a jbig2 standards-compliant encoder, the first one to appear in open source. You can get this encoder at: http://www.imperialviolet.org/jbig2.html It uses arithmetic encoding throughout. It encodes binary images losslessly with a single arithmetic coding over the full image. It also does both lossy and lossless encoding from connected components, using leptonica to generate the templates representing each cluster.
PIXA * jbAccumulateComposites ( PIXAA *pixaa, NUMA **pna, PTA **pptat )
jbAccumulateComposites() Input: pixaa (one pixa for each class) &pna (<return> number of samples used to build each composite) &ptat (<return> centroids of bordered composites) Return: pixad (accumulated sum of samples in each class), or null on error
l_int32 jbAddPage ( JBCLASSER *classer, PIX *pixs )
jbAddPage() Input: jbclasser pixs (of input page) Return: 0 if OK; 1 on error
l_int32 jbAddPageComponents ( JBCLASSER *classer, PIX *pixs, BOXA *boxas, PIXA *pixas )
jbAddPageComponents() Input: jbclasser pixs (of input page) boxas (b.b. of components for this page) pixas (components for this page) Return: 0 if OK; 1 on error Notes: (1) If there are no components on the page, we don't require input of empty boxas or pixas, although that's the typical situation.
l_int32 jbAddPages ( JBCLASSER *classer, SARRAY *safiles )
jbAddPages() Input: jbclasser safiles (of page image file names) Return: 0 if OK; 1 on error Note: (1) jbclasser makes a copy of the array of file names. (2) The caller is still responsible for destroying the input array.
JBCLASSER * jbClasserCreate ( l_int32 method, l_int32 components )
jbClasserCreate() Input: method (JB_RANKHAUS, JB_CORRELATION) components (JB_CONN_COMPS, JB_CHARACTERS, JB_WORDS) Return: jbclasser, or null on error
void jbClasserDestroy ( JBCLASSER **pclasser )
jbClasserDestroy() Input: &classer (<to be nulled>) Return: void
l_int32 jbClassifyCorrelation ( JBCLASSER *classer, BOXA *boxa, PIXA *pixas )
jbClassifyCorrelation() Input: jbclasser boxa (of new components for classification) pixas (of new components for classification) Return: 0 if OK; 1 on error
l_int32 jbClassifyRankHaus ( JBCLASSER *classer, BOXA *boxa, PIXA *pixas )
jbClassifyRankHaus() Input: jbclasser boxa (of new components for classification) pixas (of new components for classification) Return: 0 if OK; 1 on error
JBCLASSER * jbCorrelationInit ( l_int32 components, l_int32 maxwidth, l_int32 maxheight, l_float32 thresh, l_float32 weightfactor )
jbCorrelationInit() Input: components (JB_CONN_COMPS, JB_CHARACTERS, JB_WORDS) maxwidth (of component; use 0 for default) maxheight (of component; use 0 for default) thresh (value for correlation score: in [0.4 - 0.98]) weightfactor (corrects thresh for thick characters [0.0 - 1.0]) Return: jbclasser if OK; NULL on error Notes: (1) For scanned text, suggested input values are: thresh ~ [0.8 - 0.85] weightfactor ~ [0.5 - 0.6] (2) For electronically generated fonts (e.g., rasterized pdf), a very high thresh (e.g., 0.95) will not cause a significant increase in the number of classes.
JBCLASSER * jbCorrelationInitWithoutComponents ( l_int32 components, l_int32 maxwidth, l_int32 maxheight, l_float32 thresh, l_float32 weightfactor )
jbCorrelationInitWithoutComponents() Input: same as jbCorrelationInit Output: same as jbCorrelationInit Note: acts the same as jbCorrelationInit(), but the resulting object doesn't keep a list of all the components.
void jbDataDestroy ( JBDATA **pdata )
jbDataDestroy() Input: &data (<to be nulled>) Return: void
JBDATA * jbDataRead ( const char *rootname )
jbDataRead() Input: rootname (for template and data files) Return: jbdata, or NULL on error
PIXA * jbDataRender ( JBDATA *data, l_int32 debugflag )
jbDataRender() Input: jbdata debugflag (if TRUE, writes into 2 bpp pix and adds component outlines in color) Return: pixa (reconstruction of original images, using templates) or null on error
JBDATA * jbDataSave ( JBCLASSER *classer )
jbDataSave() Input: jbclasser latticew, latticeh (cell size used to store each connected component in the composite) Return: jbdata, or null on error Notes: (1) This routine stores the jbig2-type data required for generating a lossy jbig2 version of the image. It can be losslessly written to (and read from) two files. (2) It generates and stores the mosaic of templates. (3) It clones the Numa and Pta arrays, so these must all be destroyed by the caller. (4) Input 0 to use the default values for latticew and/or latticeh,
l_int32 jbDataWrite ( const char *rootout, JBDATA *jbdata )
jbDataWrite() Input: rootname (for output files; everything but the extension) jbdata Return: 0 if OK, 1 on error Notes: (1) Serialization function that writes data in jbdata to file.
l_int32 jbGetComponents ( PIX *pixs, l_int32 components, l_int32 maxwidth, l_int32 maxheight, BOXA **pboxad, PIXA **ppixad )
jbGetComponents() Input: pixs (1 bpp) components (JB_CONN_COMPS, JB_CHARACTERS, JB_WORDS) maxwidth, maxheight (of saved components; larger are discarded) &pboxa (<return> b.b. of component items) &ppixa (<return> component items) Return: 0 if OK, 1 on error
l_int32 jbGetLLCorners ( JBCLASSER *classer )
jbGetLLCorners() Input: jbclasser Return: 0 if OK, 1 on error Notes: (1) This computes the ptall field, which has the global LL corners, adjusted for each specific component, so that each component can be replaced by the template for its class and have the centroid in the template in the same position as the centroid of the original connected component. It is important that this be done properly to avoid a wavy baseline in the result. (2) It is computed here from the corresponding UL corners, where the input templates and stored instances are all bordered. This should be done after all pages have been processed. (3) For proper substitution, the templates whose LL corners are placed in these locations must be UN-bordered. This is available for a realistic jbig2 encoder, which would (1) encode each template without a border, and (2) encode the position using the LL corner (rather than the UL corner) because the difference between y-values of successive instances is typically close to zero.
l_int32 jbGetULCorners ( JBCLASSER *classer, PIX *pixs, BOXA *boxa )
jbGetULCorners() Input: jbclasser pixs (full res image) boxa (of c.c. bounding rectangles for this page) Return: 0 if OK, 1 on error Notes: (1) This computes the ptaul field, which has the global UL corners, adjusted for each specific component, so that each component can be replaced by the template for its class and have the centroid in the template in the same position as the centroid of the original connected component. It is important that this be done properly to avoid a wavy baseline in the result. (2) The array fields ptac and ptact give the centroids of those components relative to the UL corner of each component. Here, we compute the difference in each component, round to nearest integer, and correct the box->x and box->y by the appropriate integral difference. (3) The templates and stored instances are all bordered.
JBCLASSER * jbRankHausInit ( l_int32 components, l_int32 maxwidth, l_int32 maxheight, l_int32 size, l_float32 rank )
jbRankHausInit() Input: components (JB_CONN_COMPS, JB_CHARACTERS, JB_WORDS) maxwidth (of component; use 0 for default) maxheight (of component; use 0 for default) size (of square structuring element; 2, representing 2x2 sel, is necessary for reasonable accuracy of small components; combine this with rank ~ 0.97 to avoid undue class expansion) rank (rank val of match, each way; in [0.5 - 1.0]; when using size = 2, 0.97 is a reasonable value) Return: jbclasser if OK; NULL on error
PIXA * jbTemplatesFromComposites ( PIXA *pixac, NUMA *na )
jbTemplatesFromComposites() Input: pixac (one pix of composites for each class) na (number of samples used for each class composite) Return: pixad (8 bpp templates for each class), or null on error
l_int32 pixHaustest ( PIX *pix1, PIX *pix2, PIX *pix3, PIX *pix4, l_float32 delx, l_float32 dely, l_int32 maxdiffw, l_int32 maxdiffh )
pixHaustest() Input: pix1 (new pix, not dilated) pix2 (new pix, dilated) pix3 (exemplar pix, not dilated) pix4 (exemplar pix, dilated) delx (x comp of centroid difference) dely (y comp of centroid difference) maxdiffw (max width difference of pix1 and pix2) maxdiffh (max height difference of pix1 and pix2) Return: 0 (FALSE) if no match, 1 (TRUE) if the new pix is in the same class as the exemplar. Note: we check first that the two pix are roughly the same size. Only if they meet that criterion do we compare the bitmaps. The Hausdorff is a 2-way check. The centroid difference is used to align the two images to the nearest integer for each of the checks. These check that the dilated image of one contains ALL the pixels of the undilated image of the other. Checks are done in both direction. A single pixel not contained in either direction results in failure of the test.
l_int32 pixRankHaustest ( PIX *pix1, PIX *pix2, PIX *pix3, PIX *pix4, l_float32 delx, l_float32 dely, l_int32 maxdiffw, l_int32 maxdiffh, l_int32 area1, l_int32 area3, l_float32 rank, l_int32 *tab8 )
pixRankHaustest() Input: pix1 (new pix, not dilated) pix2 (new pix, dilated) pix3 (exemplar pix, not dilated) pix4 (exemplar pix, dilated) delx (x comp of centroid difference) dely (y comp of centroid difference) maxdiffw (max width difference of pix1 and pix2) maxdiffh (max height difference of pix1 and pix2) area1 (fg pixels in pix1) area3 (fg pixels in pix3) rank (rank value of test, each way) tab8 (table of pixel sums for byte) Return: 0 (FALSE) if no match, 1 (TRUE) if the new pix is in the same class as the exemplar. Note: we check first that the two pix are roughly the same size. Only if they meet that criterion do we compare the bitmaps. We convert the rank value to a number of pixels by multiplying the rank fraction by the number of pixels in the undilated image. The Hausdorff is a 2-way check. The centroid difference is used to align the two images to the nearest integer for each of the checks. The rank hausdorff checks that the dilated image of one contains the rank fraction of the pixels of the undilated image of the other. Checks are done in both direction. Failure of the test in either direction results in failure of the test.
l_int32 pixWordBoxesByDilation ( PIX *pixs, l_int32 maxdil, l_int32 minwidth, l_int32 minheight, l_int32 maxwidth, l_int32 maxheight, BOXA **pboxa, l_int32 *psize )
pixWordBoxesByDilation() Input: pixs (1 bpp; typ. at 75 to 150 ppi) maxdil (maximum dilation; 0 for default; warning if > 20) minwidth, minheight (of saved components; smaller are discarded) maxwidth, maxheight (of saved components; larger are discarded) &boxa (<return> dilated word mask) &size (<optional return> size of optimal horiz Sel) Return: 0 if OK, 1 on error Notes: (1) Returns a pruned set of word boxes. (2) See pixWordMaskByDilation().
l_int32 pixWordMaskByDilation ( PIX *pixs, l_int32 maxdil, PIX **ppixm, l_int32 *psize )
pixWordMaskByDilation() Input: pixs (1 bpp; typ. at 75 to 150 ppi) maxdil (maximum dilation; 0 for default; warning if > 20) &mask (<optional return> dilated word mask) &size (<optional return> size of optimal horiz Sel) Return: 0 if OK, 1 on error Notes: (1) This gives a crude estimate of the word masks. See pixWordBoxesByDilation() for further filtering of the word boxes. (2) For 75 to 150 ppi, the optimal dilation will be between 5 and 11. For 200 to 300 ppi, it is advisable to use a larger value for @maxdil, say between 10 and 20. Setting maxdil <= 0 results in a default dilation of 16. (3) The best size for dilating to get word masks is optionally returned.
Zakariyya Mughal <email@example.com>
This software is copyright (c) 2014 by Zakariyya Mughal.
This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.