- COPYRIGHT AND LICENSE
recogdid.c Top-level identification l_int32 recogDecode() Generate decoding arrays l_int32 recogMakeDecodingArrays() static l_int32 recogMakeDecodingArray() Dynamic programming for best path l_int32 recogRunViterbi() l_int32 recogRescoreDidResult() static PIX *recogShowPath() Create/destroy temporary DID data l_int32 recogCreateDid() l_int32 recogDestroyDid() Various helpers l_int32 recogDidExists() L_RDID *recogGetDid() static l_int32 recogGetWindowedArea() l_int32 recogSetChannelParams() static l_int32 recogTransferRchToDid() See recogbasic.c for examples of training a recognizer, which is required before it can be used for document image decoding. Gary Kopec pioneered this hidden markov approach to "Document Image Decoding" (DID) in the early 1990s. It is based on estimation using a generative model of the image generation process, and provides the most likely decoding of an image if the model is correct. Given the model, it finds the maximum a posteriori (MAP) "message" given the observed image. The model describes how to generate an image from a message, and the MAP message is derived from the observed image using Bayes' theorem. This approach can also be used to build the model, using the iterative expectation/maximization method from labelled but errorful data. In a little more detail: The model comprises three things: the ideal printed character templates, the independent bit-flip noise model, and the character setwidths. When a character is printed, the setwidth is the distance in pixels that you move forward before being able to print the next character. It is typically slightly less than the width of the character template: if too small, an extra character can be hallucinated; if too large, it will not be able to match the next character template on the line. The model assumes that the probabilities of bit flip depend only on the assignment of the pixel to background or template foreground. The multilevel templates have different bit flip probabilities for each level. Because a character image is composed of many pixels, each of which can be independently flipped, the actual probability of seeing any rendering is exceedingly small, being composed of the product of the probabilities for each pixel. The log likelihood is used both to avoid numeric underflow and, more importantly, because it results in a summation of independent pixel probabilities. That summation can be shown, in Kopec's original paper, to consist of a sum of two terms: (a) the number of fg pixels in the bit-and of the observed image with the ideal template and (b) the number of fg pixels in the template. Each has a coefficient that depends only on the bit-flip probabilities for the fg and bg. A beautiful result, and computationally simple! One nice feature of this approach is that the result of the decoding is not very sensitive to the values used for the bit flip probabilities. The procedure for finding the best decoding (MAP) for a given image goes under several names: Viterbi, dynamic programming, hidden markov model. It is called a "hidden markov model" because the templates are assumed to be printed serially and we don't know what they are -- the identity of the templates must be inferred from the observed image. The possible decodings form a dense trellis over the pixel positions, where at each pixel position you have the possibility of having any of the characters printed there (with some reference point) or having a single pixel wide space inserted there. Thus, before the trellis can be traversed, we must do the work of finding the log probability, at each pixel location, that each of the templates was printed there. Armed with those arrays of data, the dynamic programming procedure moves from left to right, one pixel at a time, recursively finding the path with the highest log probability that gets to that pixel position (and noting which template was printed to arrive there). After reaching the right side of the image, we can simply backtrack along the path, jumping over each template that lies on the highest scoring path. This best path thus only goes through a few of the pixel positions. There are two refinements to the original Kopec paper. In the first, one uses multiple, non-overlapping fg templates, each with its own bit flip probability. This makes sense, because the probability that a fg boundary pixel flips to bg is greater than that of a fg pixel not on the boundary. And the flip probability of a fg boundary pixel is smaller than that of a bg boundary pixel, which in turn is greater than that of a bg pixel not on a boundary (the latter is taken to be the true background). Then the simplest realistic multiple template model has three templates that are not background. In the second refinement, a heuristic (strict upper bound) is used iteratively in the Viterbi process to compute the log probabilities. Using the heuristic, you find the best path, and then score all nodes on that path with the actual probability, which is guaranteed to be a smaller number. You run this iteratively, rescoring just the best found path each time. After each rescoring, the path may change because the local scores have been reduced. However, the process converges rapidly, and when it doesn't change, it must be the best path because it is properly scored (even if neighboring paths are heuristically scored). The heuristic score is found column-wise by assuming that all the fg pixels in the template are on fg pixels in the image -- we just take the minimum of the number of pixels in the template and image column. This can easily give a 10-fold reduction in computation because the heuristic score can be computed much faster than the exact score. For reference, the classic paper on the approach by Kopec is: * "Document Image Decoding Using Markov Source Models", IEEE Trans. PAMI, Vol 16, No. 6, June 1994, pp 602-617. A refinement of the method for multilevel templates by Kopec is: * "Multilevel Character Templates for Document Image Decoding", Proc. SPIE 3027, Document Recognition IV, p. 168ff, 1997. Further refinements for more efficient decoding are given in these two papers, which are both stored on leptonica.org: * "Document Image Decoding using Iterated Complete Path Search", Minka, Bloomberg and Popat, Proc. SPIE Vol 4307, p. 250-258, Document Recognition and Retrieval VIII, San Jose, CA 2001. * "Document Image Decoding using Iterated Complete Path Search with Subsampled Heuristic Scoring", Bloomberg, Minka and Popat, ICDAR 2001, p. 344-349, Sept. 2001, Seattle.
l_int32 recogCreateDid ( L_RECOG *recog, PIX *pixs )
recogCreateDid() Input: recog pixs (of 1 bpp image to match) Return: 0 if OK, 1 on error
l_int32 recogDecode ( L_RECOG *recog, PIX *pixs, l_int32 nlevels, PIX **ppixdb )
recogDecode() Input: recog (with LUT's pre-computed) pixs (typically of multiple touching characters, 1 bpp) nlevels (of templates; 2 for now) &pixdb (<optional return> debug result; can be null) Return: 0 if OK, 1 on error
l_int32 recogDestroyDid ( L_RECOG *recog )
recogDestroyDid() Input: recog Return: 0 if OK, 1 on error Notes: (1) As the signature indicates, this is owned by the recog, and can only be destroyed using this function.
l_int32 recogDidExists ( L_RECOG *recog )
recogDidExists() Input: recog Return: 1 if recog->did exists; 0 if not or on error.
L_RDID * recogGetDid ( L_RECOG *recog )
recogGetDid() Input: recog Return: did (still owned by the recog), or null on error Notes: (1) This also makes sure the arrays are defined.
l_int32 recogMakeDecodingArrays ( L_RECOG *recog, PIX *pixs, l_int32 debug )
recogMakeDecodingArrays() Input: recog (with LUT's pre-computed) pixs (typically of multiple touching characters, 1 bpp) debug (1 for debug output; 0 otherwise) Return: 0 if OK, 1 on error Notes: (1) Generates the bit-and sum arrays for each character template along pixs. These are used in the dynamic programming step. (2) Previous arrays are destroyed and the new arrays are allocated. (3) The values are saved in the scoring arrays at the left edge of the template. They are used in the viterbi process at the setwidth position (which is near the RHS of the template as it is positioned on pixs) in the generated trellis.
l_int32 recogRunViterbi ( L_RECOG *recog, PIX **ppixdb )
recogRunViterbi() Input: recog (with LUT's pre-computed) &pixdb (<optional return> debug result; can be null) Return: 0 if OK, 1 on error Notes: (1) This is recursive, in that (a) we compute the score successively at all pixel positions x, (b) to compute the score at x in the trellis, for each template we look backwards to (x - setwidth) to get the score if that template were to be printed with its setwidth location at x. We save at x the template and score that maximizes the sum of the score at (x - setwidth) and the log-likelihood for the template to be printed with its LHS there.
l_int32 recogSetChannelParams ( L_RECOG *recog, l_int32 nlevels )
recogSetChannelParams() Input: recog nlevels Return: 0 if OK, 1 on error Notes: (1) This converts the independent bit-flip probabilities in the "channel" into log-likelihood coefficients on image sums. These coefficients are only defined for the non-background template levels. Thus for nlevels = 2 (one fg, one bg), only beta and gamma are used. For nlevels = 4 (three fg templates), we use beta[1-3] and gamma[1-3].
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.