|
| static l_int32 | pixCorrelationBestShift (PIX *pix1, PIX *pix2, NUMA *nasum1, NUMA *namoment1, l_int32 area2, l_int32 ycent2, l_int32 maxyshift, l_int32 *tab8, l_int32 *pdelx, l_int32 *pdely, l_float32 *pscore, l_int32 debugflag) |
| |
| static L_RCH * | rchCreate (l_int32 index, l_float32 score, char *text, l_int32 sample, l_int32 xloc, l_int32 yloc, l_int32 width) |
| |
| static L_RCHA * | rchaCreate () |
| |
| static l_int32 | transferRchToRcha (L_RCH *rch, L_RCHA *rcha) |
| |
| static void | l_showIndicatorSplitValues (NUMA *na1, NUMA *na2, NUMA *na3, NUMA *na4, NUMA *na5, NUMA *na6) |
| |
| static l_int32 | recogaSaveBestRcha (L_RECOGA *recoga, PIXA *pixa) |
| |
| static l_int32 | recogaTransferRch (L_RECOGA *recoga, L_RECOG *recog, l_int32 index) |
| |
| l_int32 | recogaIdentifyMultiple (L_RECOGA *recoga, PIX *pixs, l_int32 nitems, l_int32 minw, l_int32 minh, BOXA **pboxa, PIXA **ppixa, PIX **ppixdb, l_int32 debugsplit) |
| |
| l_int32 | recogSplitIntoCharacters (L_RECOG *recog, PIX *pixs, l_int32 minw, l_int32 minh, BOXA **pboxa, PIXA **ppixa, NUMA **pnaid, l_int32 debug) |
| |
| l_int32 | recogCorrelationBestRow (L_RECOG *recog, PIX *pixs, BOXA **pboxa, NUMA **pnascore, NUMA **pnaindex, SARRAY **psachar, l_int32 debug) |
| |
| l_int32 | recogCorrelationBestChar (L_RECOG *recog, PIX *pixs, BOX **pbox, l_float32 *pscore, l_int32 *pindex, char **pcharstr, PIX **ppixdb) |
| |
| l_int32 | recogaIdentifyPixa (L_RECOGA *recoga, PIXA *pixa, NUMA *naid, PIX **ppixdb) |
| |
| l_int32 | recogIdentifyPixa (L_RECOG *recog, PIXA *pixa, NUMA *naid, PIX **ppixdb) |
| |
| l_int32 | recogIdentifyPix (L_RECOG *recog, PIX *pixs, PIX **ppixdb) |
| |
| l_int32 | recogSkipIdentify (L_RECOG *recog) |
| |
| void | rchaDestroy (L_RCHA **prcha) |
| |
| void | rchDestroy (L_RCH **prch) |
| |
| l_int32 | rchaExtract (L_RCHA *rcha, NUMA **pnaindex, NUMA **pnascore, SARRAY **psatext, NUMA **pnasample, NUMA **pnaxloc, NUMA **pnayloc, NUMA **pnawidth) |
| |
| l_int32 | rchExtract (L_RCH *rch, l_int32 *pindex, l_float32 *pscore, char **ptext, l_int32 *psample, l_int32 *pxloc, l_int32 *pyloc, l_int32 *pwidth) |
| |
| PIX * | recogProcessToIdentify (L_RECOG *recog, PIX *pixs, l_int32 pad) |
| |
| PIX * | recogPreSplittingFilter (L_RECOG *recog, PIX *pixs, l_float32 maxasp, l_float32 minaf, l_int32 debug) |
| |
| l_int32 | recogSplittingFilter (L_RECOG *recog, PIX *pixs, l_float32 maxasp, l_float32 minaf, l_int32 *premove, l_int32 debug) |
| |
| SARRAY * | recogaExtractNumbers (L_RECOGA *recoga, BOXA *boxas, l_float32 scorethresh, l_int32 spacethresh, BOXAA **pbaa, NUMAA **pnaa) |
| |
| l_int32 | recogSetTemplateType (L_RECOG *recog, l_int32 templ_type) |
| |
| l_int32 | recogSetScaling (L_RECOG *recog, l_int32 scalew, l_int32 scaleh) |
| |
| static l_int32 pixCorrelationBestShift |
( |
PIX * |
pix1, |
|
|
PIX * |
pix2, |
|
|
NUMA * |
nasum1, |
|
|
NUMA * |
namoment1, |
|
|
l_int32 |
area2, |
|
|
l_int32 |
ycent2, |
|
|
l_int32 |
maxyshift, |
|
|
l_int32 * |
tab8, |
|
|
l_int32 * |
pdelx, |
|
|
l_int32 * |
pdely, |
|
|
l_float32 * |
pscore, |
|
|
l_int32 |
debugflag |
|
) |
| |
|
static |
pixCorrelationBestShift()
Input: pix1 (1 bpp, the unknown image; typically larger)
pix2 (1 bpp, the matching template image))
nasum1 (vertical column pixel sums for pix1)
namoment1 (vertical column first moment of pixels for pix1)
area2 (number of on pixels in pix2)
ycent2 (y component of centroid of pix2)
maxyshift (max y shift of pix2 around the location where
the centroids of pix2 and a windowed part of pix1
are vertically aligned)
tab8 (<optional> sum tab for ON pixels in byte; can be NULL)
&delx (<optional return> best x shift of pix2 relative to pix1
&dely (<optional return> best y shift of pix2 relative to pix1
&score (<optional return> maximum score found; can be NULL)
debugflag (<= 0 to skip; positive to generate output.
The integer is used to label the debug image.)
Return: 0 if OK, 1 on error
Notes: (1) This maximizes the correlation score between two 1 bpp images, one of which is typically wider. In a typical example, pix1 is a bitmap of 2 or more touching characters and pix2 is a single character template. This finds the location of pix2 that gives the largest correlation. (2) The windowed area of fg pixels and windowed first moment in the y direction are computed from the input sum and moment column arrays, @nasum1 and @namoment1 (3) This is a brute force operation. We compute the correlation at every x shift for which pix2 fits entirely within pix1, and where the centroid of pix2 is aligned, within +-maxyshift, with the centroid of a window of pix1 of the same width. The correlation is taken over the full height of pix1. This can be made more efficient.
recogaExtractNumbers()
Input: recoga
boxas (location of components)
scorethresh (min score for which we accept a component)
spacethresh (max horizontal distance allowed between digits,
use -1 for default)
&baa (<optional return> bounding boxes of identified numbers)
&naa (<optional return> scores of identified digits)
Return: sa (of identified numbers), or null on error
Notes: (1) This extracts digit data after recogaIdentifyMultiple() or lower-level identification has taken place. (2) Each string in the returned sa contains a sequence of ascii digits in a number. (3) The horizontal distance between boxes (limited by @spacethresh) is the negative of the horizontal overlap. (4) Components with a score less than @scorethresh, which may be hyphens or other small characters, will signal the end of the current sequence of digits in the number. A typical value for @scorethresh is 0.60. (5) We allow two digits to be combined if these conditions apply: (a) the first is to the left of the second (b) the second has a horizontal separation less than @spacethresh (c) the vertical overlap >= 0 (vertical separation < 0) (d) both have a score that exceeds @scorethresh (6) Each numa in the optionally returned naa contains the digit scores of a number. Each boxa in the optionally returned baa contains the bounding boxes of the digits in the number.
recogaIdentifyMultiple()
Input: recoga (with training finished)
pixs (containing typically a small number of characters)
nitems (to be identified in pix; use 0 if not known)
minw (remove components with width less than this;
use -1 for removing all noise components)
minh (remove components with height less than this;
use -1 for removing all noise components)
&boxa (<optional return> locations of identified components)
&pixa (<optional return> images of identified components)
&pixdb (<optional return> debug pix: inputs and best fits)
debugsplit (1 returns pix split debugging images)
Return: 0 if OK; 1 if nothing is found; 2 for other errors.
(Get a warning if nitems and the number found are both > 0,
but not equal to each other.)
Notes: (1) This filters the input pixa, looking for @nitems if requested. Set @nitems == 0 if you don't know how many chars to expect. (2) This bundles the filtered components into a pixa and calls recogIdentifyPixa(). If @nitems > 0, use @minw = -1 and @minh = -1 to remove all noise components. (3) Set @minw = 0 and @minh = 0 to get all noise components. Set @minw > 0 and/or @minh > 0 to retain selected noise components. All noise components are recognized as an empty string with a score of 0.0. (4) An attempt is made to order the (optionally) returned images and boxes in 2-dimensional sorted order. These can then be used to aggregate identified characters into numbers or words. One typically wants the pixa, which contains a boxa of the extracted subimages.
recogCorrelationBestChar()
Input: recog (with LUT's pre-computed)
pixs (can be of multiple touching characters, 1 bpp)
&box (<return> bounding box of best fit character)
&score (<return> correlation score)
&index (<optional return> index of class)
&charstr (<optional return> character string of class)
&pixdb (<optional return> debug pix showing input and best fit)
Return: 0 if OK, 1 on error
Notes: (1) Basic matching character splitter. Finds the best match among all templates to some region of the image. This can result in splitting the image into two parts. This is "image decoding" without dynamic programming, because we don't use a setwidth and compute the best matching score for the entire image. (2) Matching is to the average templates, without character scaling.
recogCorrelationBestRow()
Input: recog (with LUT's pre-computed)
pixs (typically of multiple touching characters, 1 bpp)
&boxa (<return> bounding boxs of best fit character)
&nascores (<optional return> correlation scores)
&naindex (<optional return> indices of classes)
&sachar (<optional return> array of character strings)
debug (1 for results written to pixadb_split)
Return: 0 if OK, 1 on error
Notes: (1) Supervises character matching for (in general) a c.c with multiple touching characters. Finds the best match greedily. Rejects small parts that are left over after splitting. (2) Matching is to the average, and without character scaling.
recogIdentifyPix()
Input: recog (with LUT's pre-computed)
pixs (of a single character, 1 bpp)
&pixdb (<optional return> debug pix showing input and best fit)
Return: 0 if OK, 1 on error
Notes: (1) Basic recognition function for a single character. (2) If L_USE_ALL, matching is attempted to every bitmap in the recog, and the identify of the best match is returned. However, if L_USE_AVERAGE, the matching is only to the averaged bitmaps, and the index of the bestsample is meaningless (0 is returned if requested). (3) The score is related to the confidence (probability of correct identification), in that a higher score is correlated with a higher probability. However, the actual relation between the correlation (score) and the probability is not known; we call this a "score" because "confidence" can be misinterpreted as an actual probability.
recogProcessToIdentify()
Input: recog (with LUT's pre-computed)
pixs (typ. single character, possibly d > 1 and uncropped)
pad (extra pixels added to left and right sides)
Return: pixd (1 bpp, clipped to foreground), or null if there
are no fg pixels or on error.
Notes: (1) This is a lightweight operation to insure that the input image is 1 bpp, properly cropped, and padded on each side. If bpp > 1, the image is thresholded.
recogSplitIntoCharacters()
Input: recog
pixs (1 bpp, contains only mostly deskewed text)
minw (remove components with width less than this;
use -1 for default removing out of band components)
minh (remove components with height less than this;
use -1 for default removing out of band components)
&boxa (<return> character bounding boxes)
&pixa (<return> character images)
&naid (<return> indices of components to identify)
debug (1 for results written to pixadb_split)
Return: 0 if OK, 1 on error or if no components are returned
Notes: (1) This can be given an image that has an arbitrary number of text characters. It does splitting of connected components based on greedy correlation matching in recogCorrelationBestRow(). The returned pixa includes the boxes from which the (possibly split) components are extracted. (2) If either @minw < 0 or @minh < 0, noise components are filtered out, and the returned @naid array is all 1. Otherwise, some noise components whose dimensions (w,h) satisfy w >= @minw and h >= @minh are allowed through, but they are identified in the returned @naid, where they are labelled by 0 to indicate that they are not to be run through identification. Retaining the noise components provides spatial information that can help applications interpret the results. (3) In addition to optional filtering of the noise, the resulting components are put in row-major (2D) order, and the smaller of overlapping components are removed if they satisfy conditions of relative size and fractional overlap. (4) Note that the spliting function uses unscaled templates and does not bother returning the class results and scores. Thes are more accurately found later using the scaled templates.
1.8.15