1 /**********************************************************************
2 * File: devanagari_processing.cpp
3 * Description: Methods to process images containing devanagari symbols,
4 * prior to classification.
5 * Author: Shobhit Saxena
6 * Created: Mon Nov 17 20:26:01 IST 2008
7 *
8 * (C) Copyright 2008, Google Inc.
9 ** Licensed under the Apache License, Version 2.0 (the "License");
10 ** you may not use this file except in compliance with the License.
11 ** You may obtain a copy of the License at
12 ** http://www.apache.org/licenses/LICENSE-2.0
13 ** Unless required by applicable law or agreed to in writing, software
14 ** distributed under the License is distributed on an "AS IS" BASIS,
15 ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 ** See the License for the specific language governing permissions and
17 ** limitations under the License.
18 *
19 **********************************************************************/
20
21 #ifdef HAVE_CONFIG_H
22 # include "config_auto.h"
23 #endif
24
25 #include "devanagari_processing.h"
26
27 #include "debugpixa.h"
28 #include "statistc.h"
29 #include "tordmain.h"
30
31 #include <allheaders.h>
32
33 namespace tesseract {
34
35 // Flags controlling the debugging information for shiro-rekha splitting
36 // strategies.
37 INT_VAR(devanagari_split_debuglevel, 0, "Debug level for split shiro-rekha process.");
38
39 BOOL_VAR(devanagari_split_debugimage, 0,
40 "Whether to create a debug image for split shiro-rekha process.");
41
ShiroRekhaSplitter()42 ShiroRekhaSplitter::ShiroRekhaSplitter() {
43 orig_pix_ = nullptr;
44 segmentation_block_list_ = nullptr;
45 splitted_image_ = nullptr;
46 global_xheight_ = kUnspecifiedXheight;
47 perform_close_ = false;
48 debug_image_ = nullptr;
49 pageseg_split_strategy_ = NO_SPLIT;
50 ocr_split_strategy_ = NO_SPLIT;
51 }
52
~ShiroRekhaSplitter()53 ShiroRekhaSplitter::~ShiroRekhaSplitter() {
54 Clear();
55 }
56
Clear()57 void ShiroRekhaSplitter::Clear() {
58 orig_pix_.destroy();
59 splitted_image_.destroy();
60 pageseg_split_strategy_ = NO_SPLIT;
61 ocr_split_strategy_ = NO_SPLIT;
62 debug_image_.destroy();
63 segmentation_block_list_ = nullptr;
64 global_xheight_ = kUnspecifiedXheight;
65 perform_close_ = false;
66 }
67
68 // On setting the input image, a clone of it is owned by this class.
set_orig_pix(Image pix)69 void ShiroRekhaSplitter::set_orig_pix(Image pix) {
70 if (orig_pix_) {
71 orig_pix_.destroy();
72 }
73 orig_pix_ = pix.clone();
74 }
75
76 // Top-level method to perform splitting based on current settings.
77 // Returns true if a split was actually performed.
78 // split_for_pageseg should be true if the splitting is being done prior to
79 // page segmentation. This mode uses the flag
80 // pageseg_devanagari_split_strategy to determine the splitting strategy.
Split(bool split_for_pageseg,DebugPixa * pixa_debug)81 bool ShiroRekhaSplitter::Split(bool split_for_pageseg, DebugPixa *pixa_debug) {
82 SplitStrategy split_strategy = split_for_pageseg ? pageseg_split_strategy_ : ocr_split_strategy_;
83 if (split_strategy == NO_SPLIT) {
84 return false; // Nothing to do.
85 }
86 ASSERT_HOST(split_strategy == MINIMAL_SPLIT || split_strategy == MAXIMAL_SPLIT);
87 ASSERT_HOST(orig_pix_);
88 if (devanagari_split_debuglevel > 0) {
89 tprintf("Splitting shiro-rekha ...\n");
90 tprintf("Split strategy = %s\n", split_strategy == MINIMAL_SPLIT ? "Minimal" : "Maximal");
91 tprintf("Initial pageseg available = %s\n", segmentation_block_list_ ? "yes" : "no");
92 }
93 // Create a copy of original image to store the splitting output.
94 splitted_image_.destroy();
95 splitted_image_ = orig_pix_.copy();
96
97 // Initialize debug image if required.
98 if (devanagari_split_debugimage) {
99 debug_image_.destroy();
100 debug_image_ = pixConvertTo32(orig_pix_);
101 }
102
103 // Determine all connected components in the input image. A close operation
104 // may be required prior to this, depending on the current settings.
105 Image pix_for_ccs = orig_pix_.clone();
106 if (perform_close_ && global_xheight_ != kUnspecifiedXheight && !segmentation_block_list_) {
107 if (devanagari_split_debuglevel > 0) {
108 tprintf("Performing a global close operation..\n");
109 }
110 // A global measure is available for xheight, but no local information
111 // exists.
112 pix_for_ccs.destroy();
113 pix_for_ccs = orig_pix_.copy();
114 PerformClose(pix_for_ccs, global_xheight_);
115 }
116 Pixa *ccs;
117 Boxa *tmp_boxa = pixConnComp(pix_for_ccs, &ccs, 8);
118 boxaDestroy(&tmp_boxa);
119 pix_for_ccs.destroy();
120
121 // Iterate over all connected components. Get their bounding boxes and clip
122 // out the image regions corresponding to these boxes from the original image.
123 // Conditionally run splitting on each of them.
124 Boxa *regions_to_clear = boxaCreate(0);
125 int num_ccs = 0;
126 if (ccs != nullptr) {
127 num_ccs = pixaGetCount(ccs);
128 }
129 for (int i = 0; i < num_ccs; ++i) {
130 Box *box = ccs->boxa->box[i];
131 Image word_pix = pixClipRectangle(orig_pix_, box, nullptr);
132 ASSERT_HOST(word_pix);
133 int xheight = GetXheightForCC(box);
134 if (xheight == kUnspecifiedXheight && segmentation_block_list_ && devanagari_split_debugimage) {
135 pixRenderBoxArb(debug_image_, box, 1, 255, 0, 0);
136 }
137 // If some xheight measure is available, attempt to pre-eliminate small
138 // blobs from the shiro-rekha process. This is primarily to save the CCs
139 // corresponding to punctuation marks/small dots etc which are part of
140 // larger graphemes.
141 if (xheight == kUnspecifiedXheight || (box->w > xheight / 3 && box->h > xheight / 2)) {
142 SplitWordShiroRekha(split_strategy, word_pix, xheight, box->x, box->y, regions_to_clear);
143 } else if (devanagari_split_debuglevel > 0) {
144 tprintf("CC dropped from splitting: %d,%d (%d, %d)\n", box->x, box->y, box->w, box->h);
145 }
146 word_pix.destroy();
147 }
148 // Actually clear the boxes now.
149 for (int i = 0; i < boxaGetCount(regions_to_clear); ++i) {
150 Box *box = boxaGetBox(regions_to_clear, i, L_CLONE);
151 pixClearInRect(splitted_image_, box);
152 boxDestroy(&box);
153 }
154 boxaDestroy(®ions_to_clear);
155 pixaDestroy(&ccs);
156 if (devanagari_split_debugimage && pixa_debug != nullptr) {
157 pixa_debug->AddPix(debug_image_, split_for_pageseg ? "pageseg_split" : "ocr_split");
158 }
159 return true;
160 }
161
162 // Method to perform a close operation on the input image. The xheight
163 // estimate decides the size of sel used.
PerformClose(Image pix,int xheight_estimate)164 void ShiroRekhaSplitter::PerformClose(Image pix, int xheight_estimate) {
165 pixCloseBrick(pix, pix, xheight_estimate / 8, xheight_estimate / 3);
166 }
167
168 // This method resolves the cc bbox to a particular row and returns the row's
169 // xheight.
GetXheightForCC(Box * cc_bbox)170 int ShiroRekhaSplitter::GetXheightForCC(Box *cc_bbox) {
171 if (!segmentation_block_list_) {
172 return global_xheight_;
173 }
174 // Compute the box coordinates in Tesseract's coordinate system.
175 TBOX bbox(cc_bbox->x, pixGetHeight(orig_pix_) - cc_bbox->y - cc_bbox->h - 1,
176 cc_bbox->x + cc_bbox->w, pixGetHeight(orig_pix_) - cc_bbox->y - 1);
177 // Iterate over all blocks.
178 BLOCK_IT block_it(segmentation_block_list_);
179 for (block_it.mark_cycle_pt(); !block_it.cycled_list(); block_it.forward()) {
180 BLOCK *block = block_it.data();
181 // Iterate over all rows in the block.
182 ROW_IT row_it(block->row_list());
183 for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
184 ROW *row = row_it.data();
185 if (!row->bounding_box().major_overlap(bbox)) {
186 continue;
187 }
188 // Row could be skewed, warped, etc. Use the position of the box to
189 // determine the baseline position of the row for that x-coordinate.
190 // Create a square TBOX whose baseline's mid-point lies at this point
191 // and side is row's xheight. Take the overlap of this box with the input
192 // box and check if it is a 'major overlap'. If so, this box lies in this
193 // row. In that case, return the xheight for this row.
194 float box_middle = 0.5 * (bbox.left() + bbox.right());
195 int baseline = static_cast<int>(row->base_line(box_middle) + 0.5);
196 TBOX test_box(box_middle - row->x_height() / 2, baseline, box_middle + row->x_height() / 2,
197 static_cast<int>(baseline + row->x_height()));
198 // Compute overlap. If it is is a major overlap, this is the right row.
199 if (bbox.major_overlap(test_box)) {
200 return row->x_height();
201 }
202 }
203 }
204 // No row found for this bbox.
205 return kUnspecifiedXheight;
206 }
207
208 // Returns a list of regions (boxes) which should be cleared in the original
209 // image so as to perform shiro-rekha splitting. Pix is assumed to carry one
210 // (or less) word only. Xheight measure could be the global estimate, the row
211 // estimate, or unspecified. If unspecified, over splitting may occur, since a
212 // conservative estimate of stroke width along with an associated multiplier
213 // is used in its place. It is advisable to have a specified xheight when
214 // splitting for classification/training.
215 // A vertical projection histogram of all the on-pixels in the input pix is
216 // computed. The maxima of this histogram is regarded as an approximate location
217 // of the shiro-rekha. By descending on the maxima's peak on both sides,
218 // stroke width of shiro-rekha is estimated.
219 // A horizontal projection histogram is computed for a sub-image of the input
220 // image, which extends from just below the shiro-rekha down to a certain
221 // leeway. The leeway depends on the input xheight, if provided, else a
222 // conservative multiplier on approximate stroke width is used (which may lead
223 // to over-splitting).
SplitWordShiroRekha(SplitStrategy split_strategy,Image pix,int xheight,int word_left,int word_top,Boxa * regions_to_clear)224 void ShiroRekhaSplitter::SplitWordShiroRekha(SplitStrategy split_strategy, Image pix, int xheight,
225 int word_left, int word_top, Boxa *regions_to_clear) {
226 if (split_strategy == NO_SPLIT) {
227 return;
228 }
229 int width = pixGetWidth(pix);
230 int height = pixGetHeight(pix);
231 // Statistically determine the yextents of the shiro-rekha.
232 int shirorekha_top, shirorekha_bottom, shirorekha_ylevel;
233 GetShiroRekhaYExtents(pix, &shirorekha_top, &shirorekha_bottom, &shirorekha_ylevel);
234 // Since the shiro rekha is also a stroke, its width is equal to the stroke
235 // width.
236 int stroke_width = shirorekha_bottom - shirorekha_top + 1;
237
238 // Some safeguards to protect CCs we do not want to be split.
239 // These are particularly useful when the word wasn't eliminated earlier
240 // because xheight information was unavailable.
241 if (shirorekha_ylevel > height / 2) {
242 // Shirorekha shouldn't be in the bottom half of the word.
243 if (devanagari_split_debuglevel > 0) {
244 tprintf("Skipping splitting CC at (%d, %d): shirorekha in lower half..\n", word_left,
245 word_top);
246 }
247 return;
248 }
249 if (stroke_width > height / 3) {
250 // Even the boldest of fonts shouldn't do this.
251 if (devanagari_split_debuglevel > 0) {
252 tprintf("Skipping splitting CC at (%d, %d): stroke width too huge..\n", word_left, word_top);
253 }
254 return;
255 }
256
257 // Clear the ascender and descender regions of the word.
258 // Obtain a vertical projection histogram for the resulting image.
259 Box *box_to_clear = boxCreate(0, shirorekha_top - stroke_width / 3, width, 5 * stroke_width / 3);
260 Image word_in_xheight = pix.copy();
261 pixClearInRect(word_in_xheight, box_to_clear);
262 // Also clear any pixels which are below shirorekha_bottom + some leeway.
263 // The leeway is set to xheight if the information is available, else it is a
264 // multiplier applied to the stroke width.
265 int leeway_to_keep = stroke_width * 3;
266 if (xheight != kUnspecifiedXheight) {
267 // This is because the xheight-region typically includes the shiro-rekha
268 // inside it, i.e., the top of the xheight range corresponds to the top of
269 // shiro-rekha.
270 leeway_to_keep = xheight - stroke_width;
271 }
272 box_to_clear->y = shirorekha_bottom + leeway_to_keep;
273 box_to_clear->h = height - box_to_clear->y;
274 pixClearInRect(word_in_xheight, box_to_clear);
275 boxDestroy(&box_to_clear);
276
277 PixelHistogram vert_hist;
278 vert_hist.ConstructVerticalCountHist(word_in_xheight);
279 word_in_xheight.destroy();
280
281 // If the number of black pixel in any column of the image is less than a
282 // fraction of the stroke width, treat it as noise / a stray mark. Perform
283 // these changes inside the vert_hist data itself, as that is used later on as
284 // a bit vector for the final split decision at every column.
285 for (int i = 0; i < width; ++i) {
286 if (vert_hist.hist()[i] <= stroke_width / 4) {
287 vert_hist.hist()[i] = 0;
288 } else {
289 vert_hist.hist()[i] = 1;
290 }
291 }
292 // In order to split the line at any point, we make sure that the width of the
293 // gap is at least half the stroke width.
294 int i = 0;
295 int cur_component_width = 0;
296 while (i < width) {
297 if (!vert_hist.hist()[i]) {
298 int j = 0;
299 while (i + j < width && !vert_hist.hist()[i + j]) {
300 ++j;
301 }
302 if (j >= stroke_width / 2 && cur_component_width >= stroke_width / 2) {
303 // Perform a shiro-rekha split. The intervening region lies from i to
304 // i+j-1.
305 // A minimal single-pixel split makes the estimation of intra- and
306 // inter-word spacing easier during page layout analysis,
307 // whereas a maximal split may be needed for OCR, depending on
308 // how the engine was trained.
309 bool minimal_split = (split_strategy == MINIMAL_SPLIT);
310 int split_width = minimal_split ? 1 : j;
311 int split_left = minimal_split ? i + (j / 2) - (split_width / 2) : i;
312 if (!minimal_split || (i != 0 && i + j != width)) {
313 Box *box_to_clear =
314 boxCreate(word_left + split_left, word_top + shirorekha_top - stroke_width / 3,
315 split_width, 5 * stroke_width / 3);
316 if (box_to_clear) {
317 boxaAddBox(regions_to_clear, box_to_clear, L_CLONE);
318 // Mark this in the debug image if needed.
319 if (devanagari_split_debugimage) {
320 pixRenderBoxArb(debug_image_, box_to_clear, 1, 128, 255, 128);
321 }
322 boxDestroy(&box_to_clear);
323 cur_component_width = 0;
324 }
325 }
326 }
327 i += j;
328 } else {
329 ++i;
330 ++cur_component_width;
331 }
332 }
333 }
334
335 // Refreshes the words in the segmentation block list by using blobs in the
336 // input block list.
337 // The segmentation block list must be set.
RefreshSegmentationWithNewBlobs(C_BLOB_LIST * new_blobs)338 void ShiroRekhaSplitter::RefreshSegmentationWithNewBlobs(C_BLOB_LIST *new_blobs) {
339 // The segmentation block list must have been specified.
340 ASSERT_HOST(segmentation_block_list_);
341 if (devanagari_split_debuglevel > 0) {
342 tprintf("Before refreshing blobs:\n");
343 PrintSegmentationStats(segmentation_block_list_);
344 tprintf("New Blobs found: %d\n", new_blobs->length());
345 }
346
347 C_BLOB_LIST not_found_blobs;
348 RefreshWordBlobsFromNewBlobs(
349 segmentation_block_list_, new_blobs,
350 ((devanagari_split_debugimage && debug_image_) ? ¬_found_blobs : nullptr));
351
352 if (devanagari_split_debuglevel > 0) {
353 tprintf("After refreshing blobs:\n");
354 PrintSegmentationStats(segmentation_block_list_);
355 }
356 if (devanagari_split_debugimage && debug_image_) {
357 // Plot out the original blobs for which no match was found in the new
358 // all_blobs list.
359 C_BLOB_IT not_found_it(¬_found_blobs);
360 for (not_found_it.mark_cycle_pt(); !not_found_it.cycled_list(); not_found_it.forward()) {
361 C_BLOB *not_found = not_found_it.data();
362 TBOX not_found_box = not_found->bounding_box();
363 Box *box_to_plot = GetBoxForTBOX(not_found_box);
364 pixRenderBoxArb(debug_image_, box_to_plot, 1, 255, 0, 255);
365 boxDestroy(&box_to_plot);
366 }
367
368 // Plot out the blobs unused from all blobs.
369 C_BLOB_IT all_blobs_it(new_blobs);
370 for (all_blobs_it.mark_cycle_pt(); !all_blobs_it.cycled_list(); all_blobs_it.forward()) {
371 C_BLOB *a_blob = all_blobs_it.data();
372 Box *box_to_plot = GetBoxForTBOX(a_blob->bounding_box());
373 pixRenderBoxArb(debug_image_, box_to_plot, 3, 0, 127, 0);
374 boxDestroy(&box_to_plot);
375 }
376 }
377 }
378
379 // Returns a new box object for the corresponding TBOX, based on the original
380 // image's coordinate system.
GetBoxForTBOX(const TBOX & tbox) const381 Box *ShiroRekhaSplitter::GetBoxForTBOX(const TBOX &tbox) const {
382 return boxCreate(tbox.left(), pixGetHeight(orig_pix_) - tbox.top() - 1, tbox.width(),
383 tbox.height());
384 }
385
386 // This method returns the computed mode-height of blobs in the pix.
387 // It also prunes very small blobs from calculation.
GetModeHeight(Image pix)388 int ShiroRekhaSplitter::GetModeHeight(Image pix) {
389 Boxa *boxa = pixConnComp(pix, nullptr, 8);
390 STATS heights(0, pixGetHeight(pix));
391 heights.clear();
392 for (int i = 0; i < boxaGetCount(boxa); ++i) {
393 Box *box = boxaGetBox(boxa, i, L_CLONE);
394 if (box->h >= 3 || box->w >= 3) {
395 heights.add(box->h, 1);
396 }
397 boxDestroy(&box);
398 }
399 boxaDestroy(&boxa);
400 return heights.mode();
401 }
402
403 // This method returns y-extents of the shiro-rekha computed from the input
404 // word image.
GetShiroRekhaYExtents(Image word_pix,int * shirorekha_top,int * shirorekha_bottom,int * shirorekha_ylevel)405 void ShiroRekhaSplitter::GetShiroRekhaYExtents(Image word_pix, int *shirorekha_top,
406 int *shirorekha_bottom, int *shirorekha_ylevel) {
407 // Compute a histogram from projecting the word on a vertical line.
408 PixelHistogram hist_horiz;
409 hist_horiz.ConstructHorizontalCountHist(word_pix);
410 // Get the ylevel where the top-line exists. This is basically the global
411 // maxima in the horizontal histogram.
412 int topline_onpixel_count = 0;
413 int topline_ylevel = hist_horiz.GetHistogramMaximum(&topline_onpixel_count);
414
415 // Get the upper and lower extents of the shiro rekha.
416 int thresh = (topline_onpixel_count * 70) / 100;
417 int ulimit = topline_ylevel;
418 int llimit = topline_ylevel;
419 while (ulimit > 0 && hist_horiz.hist()[ulimit] >= thresh) {
420 --ulimit;
421 }
422 while (llimit < pixGetHeight(word_pix) && hist_horiz.hist()[llimit] >= thresh) {
423 ++llimit;
424 }
425
426 if (shirorekha_top) {
427 *shirorekha_top = ulimit;
428 }
429 if (shirorekha_bottom) {
430 *shirorekha_bottom = llimit;
431 }
432 if (shirorekha_ylevel) {
433 *shirorekha_ylevel = topline_ylevel;
434 }
435 }
436
437 // This method returns the global-maxima for the histogram. The frequency of
438 // the global maxima is returned in count, if specified.
GetHistogramMaximum(int * count) const439 int PixelHistogram::GetHistogramMaximum(int *count) const {
440 int best_value = 0;
441 for (int i = 0; i < length_; ++i) {
442 if (hist_[i] > hist_[best_value]) {
443 best_value = i;
444 }
445 }
446 if (count) {
447 *count = hist_[best_value];
448 }
449 return best_value;
450 }
451
452 // Methods to construct histograms from images.
ConstructVerticalCountHist(Image pix)453 void PixelHistogram::ConstructVerticalCountHist(Image pix) {
454 Clear();
455 int width = pixGetWidth(pix);
456 int height = pixGetHeight(pix);
457 hist_ = new int[width];
458 length_ = width;
459 int wpl = pixGetWpl(pix);
460 l_uint32 *data = pixGetData(pix);
461 for (int i = 0; i < width; ++i) {
462 hist_[i] = 0;
463 }
464 for (int i = 0; i < height; ++i) {
465 l_uint32 *line = data + i * wpl;
466 for (int j = 0; j < width; ++j) {
467 if (GET_DATA_BIT(line, j)) {
468 ++(hist_[j]);
469 }
470 }
471 }
472 }
473
ConstructHorizontalCountHist(Image pix)474 void PixelHistogram::ConstructHorizontalCountHist(Image pix) {
475 Clear();
476 Numa *counts = pixCountPixelsByRow(pix, nullptr);
477 length_ = numaGetCount(counts);
478 hist_ = new int[length_];
479 for (int i = 0; i < length_; ++i) {
480 l_int32 val = 0;
481 numaGetIValue(counts, i, &val);
482 hist_[i] = val;
483 }
484 numaDestroy(&counts);
485 }
486
487 } // namespace tesseract.
488