Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/textord/ccnontextdetect.cpp @ 2:b50eed0cc0ef upstream

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
ADD: MuPDF v1.26.7: the MuPDF source as downloaded by a default build of PyMuPDF 1.26.4. The directory name has changed: no version number in the expanded directory now.
author Franz Glasner <fzglas.hg@dom66.de>
date Mon, 15 Sep 2025 11:43:07 +0200
parents
children
comparison
equal deleted inserted replaced
1:1d09e1dec1d9 2:b50eed0cc0ef
1 ///////////////////////////////////////////////////////////////////////
2 // File: ccnontextdetect.cpp
3 // Description: Connected-Component-based photo (non-text) detection.
4 // Author: rays@google.com (Ray Smith)
5 //
6 // Copyright 2011 Google Inc. All Rights Reserved.
7 // Licensed under the Apache License, Version 2.0 (the "License");
8 // you may not use this file except in compliance with the License.
9 // You may obtain a copy of the License at
10 // http://www.apache.org/licenses/LICENSE-2.0
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS,
13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 // See the License for the specific language governing permissions and
15 // limitations under the License.
16 //
17 ///////////////////////////////////////////////////////////////////////
18
19 #ifdef HAVE_CONFIG_H
20 # include "config_auto.h"
21 #endif
22
23 #include "ccnontextdetect.h"
24 #include "helpers.h" // for IntCastRounded
25 #include "imagefind.h"
26 #include "strokewidth.h"
27
28 namespace tesseract {
29
30 // Max number of neighbour small objects per squared gridsize before a grid
31 // cell becomes image.
32 const double kMaxSmallNeighboursPerPix = 1.0 / 32;
33 // Max number of small blobs a large blob may overlap before it is rejected
34 // and determined to be image.
35 const int kMaxLargeOverlapsWithSmall = 3;
36 // Max number of small blobs a medium blob may overlap before it is rejected
37 // and determined to be image. Larger than for large blobs as medium blobs
38 // may be complex Chinese characters. Very large Chinese characters are going
39 // to overlap more medium blobs than small.
40 const int kMaxMediumOverlapsWithSmall = 12;
41 // Max number of normal blobs a large blob may overlap before it is rejected
42 // and determined to be image. This is set higher to allow for drop caps, which
43 // may overlap a lot of good text blobs.
44 const int kMaxLargeOverlapsWithMedium = 12;
45 // Multiplier of original noise_count used to test for the case of spreading
46 // noise beyond where it should really be.
47 const int kOriginalNoiseMultiple = 8;
48 // Pixel padding for noise blobs when rendering on the image
49 // mask to encourage them to join together. Make it too big and images
50 // will fatten out too much and have to be clipped to text.
51 const int kNoisePadding = 4;
52 // Fraction of max_noise_count_ to be added to the noise count if there is
53 // photo mask in the background.
54 const double kPhotoOffsetFraction = 0.375;
55 // Min ratio of perimeter^2/16area for a "good" blob in estimating noise
56 // density. Good blobs are supposed to be highly likely real text.
57 // We consider a square to have unit ratio, where A=(p/4)^2, hence the factor
58 // of 16. Digital circles are weird and have a minimum ratio of pi/64, not
59 // the 1/(4pi) that you would expect.
60 const double kMinGoodTextPARatio = 1.5;
61
62 CCNonTextDetect::CCNonTextDetect(int gridsize, const ICOORD &bleft, const ICOORD &tright)
63 : BlobGrid(gridsize, bleft, tright)
64 , max_noise_count_(static_cast<int>(kMaxSmallNeighboursPerPix * gridsize * gridsize))
65 , noise_density_(nullptr) {
66 // TODO(rays) break max_noise_count_ out into an area-proportional
67 // value, as now plus an additive constant for the number of text blobs
68 // in the 3x3 neighbourhood - maybe 9.
69 }
70
71 CCNonTextDetect::~CCNonTextDetect() {
72 delete noise_density_;
73 }
74
75 // Creates and returns a Pix with the same resolution as the original
76 // in which 1 (black) pixels represent likely non text (photo, line drawing)
77 // areas of the page, deleting from the blob_block the blobs that were
78 // determined to be non-text.
79 // The photo_map is used to bias the decision towards non-text, rather than
80 // supplying definite decision.
81 // The blob_block is the usual result of connected component analysis,
82 // holding the detected blobs.
83 // The returned Pix should be PixDestroyed after use.
84 Image CCNonTextDetect::ComputeNonTextMask(bool debug, Image photo_map, TO_BLOCK *blob_block) {
85 // Insert the smallest blobs into the grid.
86 InsertBlobList(&blob_block->small_blobs);
87 InsertBlobList(&blob_block->noise_blobs);
88 // Add the medium blobs that don't have a good strokewidth neighbour.
89 // Those that do go into good_grid as an antidote to spreading beyond the
90 // real reaches of a noise region.
91 BlobGrid good_grid(gridsize(), bleft(), tright());
92 BLOBNBOX_IT blob_it(&blob_block->blobs);
93 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
94 BLOBNBOX *blob = blob_it.data();
95 double perimeter_area_ratio = blob->cblob()->perimeter() / 4.0;
96 perimeter_area_ratio *= perimeter_area_ratio / blob->enclosed_area();
97 if (blob->GoodTextBlob() == 0 || perimeter_area_ratio < kMinGoodTextPARatio) {
98 InsertBBox(true, true, blob);
99 } else {
100 good_grid.InsertBBox(true, true, blob);
101 }
102 }
103 noise_density_ = ComputeNoiseDensity(debug, photo_map, &good_grid);
104 good_grid.Clear(); // Not needed any more.
105 Image pix = noise_density_->ThresholdToPix(max_noise_count_);
106 if (debug) {
107 pixWrite("junknoisemask.png", pix, IFF_PNG);
108 }
109 ScrollView *win = nullptr;
110 #ifndef GRAPHICS_DISABLED
111 if (debug) {
112 win = MakeWindow(0, 400, "Photo Mask Blobs");
113 }
114 #endif // !GRAPHICS_DISABLED
115 // Large and medium blobs are not text if they overlap with "a lot" of small
116 // blobs.
117 MarkAndDeleteNonTextBlobs(&blob_block->large_blobs, kMaxLargeOverlapsWithSmall, win,
118 ScrollView::DARK_GREEN, pix);
119 MarkAndDeleteNonTextBlobs(&blob_block->blobs, kMaxMediumOverlapsWithSmall, win, ScrollView::WHITE,
120 pix);
121 // Clear the grid of small blobs and insert the medium blobs.
122 Clear();
123 InsertBlobList(&blob_block->blobs);
124 MarkAndDeleteNonTextBlobs(&blob_block->large_blobs, kMaxLargeOverlapsWithMedium, win,
125 ScrollView::DARK_GREEN, pix);
126 // Clear again before we start deleting the blobs in the grid.
127 Clear();
128 MarkAndDeleteNonTextBlobs(&blob_block->noise_blobs, -1, win, ScrollView::CORAL, pix);
129 MarkAndDeleteNonTextBlobs(&blob_block->small_blobs, -1, win, ScrollView::GOLDENROD, pix);
130 MarkAndDeleteNonTextBlobs(&blob_block->blobs, -1, win, ScrollView::WHITE, pix);
131 if (debug) {
132 #ifndef GRAPHICS_DISABLED
133 win->Update();
134 #endif // !GRAPHICS_DISABLED
135 pixWrite("junkccphotomask.png", pix, IFF_PNG);
136 #ifndef GRAPHICS_DISABLED
137 win->AwaitEvent(SVET_DESTROY);
138 delete win;
139 #endif // !GRAPHICS_DISABLED
140 }
141 return pix;
142 }
143
144 // Computes and returns the noise_density IntGrid, at the same gridsize as
145 // this by summing the number of small elements in a 3x3 neighbourhood of
146 // each grid cell. good_grid is filled with blobs that are considered most
147 // likely good text, and this is filled with small and medium blobs that are
148 // more likely non-text.
149 // The photo_map is used to bias the decision towards non-text, rather than
150 // supplying definite decision.
151 IntGrid *CCNonTextDetect::ComputeNoiseDensity(bool debug, Image photo_map, BlobGrid *good_grid) {
152 IntGrid *noise_counts = CountCellElements();
153 IntGrid *noise_density = noise_counts->NeighbourhoodSum();
154 IntGrid *good_counts = good_grid->CountCellElements();
155 // Now increase noise density in photo areas, to bias the decision and
156 // minimize hallucinated text on image, but trim the noise_density where
157 // there are good blobs and the original count is low in non-photo areas,
158 // indicating that most of the result came from neighbouring cells.
159 int height = pixGetHeight(photo_map);
160 int photo_offset = IntCastRounded(max_noise_count_ * kPhotoOffsetFraction);
161 for (int y = 0; y < gridheight(); ++y) {
162 for (int x = 0; x < gridwidth(); ++x) {
163 int noise = noise_density->GridCellValue(x, y);
164 if (max_noise_count_ < noise + photo_offset && noise <= max_noise_count_) {
165 // Test for photo.
166 int left = x * gridsize();
167 int right = left + gridsize();
168 int bottom = height - y * gridsize();
169 int top = bottom - gridsize();
170 if (ImageFind::BoundsWithinRect(photo_map, &left, &top, &right, &bottom)) {
171 noise_density->SetGridCell(x, y, noise + photo_offset);
172 }
173 }
174 if (debug && noise > max_noise_count_ && good_counts->GridCellValue(x, y) > 0) {
175 tprintf("At %d, %d, noise = %d, good=%d, orig=%d, thr=%d\n", x * gridsize(), y * gridsize(),
176 noise_density->GridCellValue(x, y), good_counts->GridCellValue(x, y),
177 noise_counts->GridCellValue(x, y), max_noise_count_);
178 }
179 if (noise > max_noise_count_ && good_counts->GridCellValue(x, y) > 0 &&
180 noise_counts->GridCellValue(x, y) * kOriginalNoiseMultiple <= max_noise_count_) {
181 noise_density->SetGridCell(x, y, 0);
182 }
183 }
184 }
185 delete noise_counts;
186 delete good_counts;
187 return noise_density;
188 }
189
190 // Helper to expand a box in one of the 4 directions by the given pad,
191 // provided it does not expand into any cell with a zero noise density.
192 // If that is not possible, try expanding all round by a small constant.
193 static TBOX AttemptBoxExpansion(const TBOX &box, const IntGrid &noise_density, int pad) {
194 TBOX expanded_box(box);
195 expanded_box.set_right(box.right() + pad);
196 if (!noise_density.AnyZeroInRect(expanded_box)) {
197 return expanded_box;
198 }
199 expanded_box = box;
200 expanded_box.set_left(box.left() - pad);
201 if (!noise_density.AnyZeroInRect(expanded_box)) {
202 return expanded_box;
203 }
204 expanded_box = box;
205 expanded_box.set_top(box.top() + pad);
206 if (!noise_density.AnyZeroInRect(expanded_box)) {
207 return expanded_box;
208 }
209 expanded_box = box;
210 expanded_box.set_bottom(box.bottom() + pad);
211 if (!noise_density.AnyZeroInRect(expanded_box)) {
212 return expanded_box;
213 }
214 expanded_box = box;
215 expanded_box.pad(kNoisePadding, kNoisePadding);
216 if (!noise_density.AnyZeroInRect(expanded_box)) {
217 return expanded_box;
218 }
219 return box;
220 }
221
222 // Tests each blob in the list to see if it is certain non-text using 2
223 // conditions:
224 // 1. blob overlaps a cell with high value in noise_density_ (previously set
225 // by ComputeNoiseDensity).
226 // OR 2. The blob overlaps more than max_blob_overlaps in *this grid. This
227 // condition is disabled with max_blob_overlaps == -1.
228 // If it does, the blob is declared non-text, and is used to mark up the
229 // nontext_mask. Such blobs are fully deleted, and non-noise blobs have their
230 // neighbours reset, as they may now point to deleted data.
231 // WARNING: The blobs list blobs may be in the *this grid, but they are
232 // not removed. If any deleted blobs might be in *this, then this must be
233 // Clear()ed immediately after MarkAndDeleteNonTextBlobs is called.
234 // If the win is not nullptr, deleted blobs are drawn on it in red, and kept
235 // blobs are drawn on it in ok_color.
236 void CCNonTextDetect::MarkAndDeleteNonTextBlobs(BLOBNBOX_LIST *blobs, int max_blob_overlaps,
237 ScrollView *win, ScrollView::Color ok_color,
238 Image nontext_mask) {
239 int imageheight = tright().y() - bleft().x();
240 BLOBNBOX_IT blob_it(blobs);
241 BLOBNBOX_LIST dead_blobs;
242 BLOBNBOX_IT dead_it(&dead_blobs);
243 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
244 BLOBNBOX *blob = blob_it.data();
245 TBOX box = blob->bounding_box();
246 if (!noise_density_->RectMostlyOverThreshold(box, max_noise_count_) &&
247 (max_blob_overlaps < 0 || !BlobOverlapsTooMuch(blob, max_blob_overlaps))) {
248 blob->ClearNeighbours();
249 #ifndef GRAPHICS_DISABLED
250 if (win != nullptr) {
251 blob->plot(win, ok_color, ok_color);
252 }
253 #endif // !GRAPHICS_DISABLED
254 } else {
255 if (noise_density_->AnyZeroInRect(box)) {
256 // There is a danger that the bounding box may overlap real text, so
257 // we need to render the outline.
258 Image blob_pix = blob->cblob()->render_outline();
259 pixRasterop(nontext_mask, box.left(), imageheight - box.top(), box.width(), box.height(),
260 PIX_SRC | PIX_DST, blob_pix, 0, 0);
261 blob_pix.destroy();
262 } else {
263 if (box.area() < gridsize() * gridsize()) {
264 // It is a really bad idea to make lots of small components in the
265 // photo mask, so try to join it to a bigger area by expanding the
266 // box in a way that does not touch any zero noise density cell.
267 box = AttemptBoxExpansion(box, *noise_density_, gridsize());
268 }
269 // All overlapped cells are non-zero, so just mark the rectangle.
270 pixRasterop(nontext_mask, box.left(), imageheight - box.top(), box.width(), box.height(),
271 PIX_SET, nullptr, 0, 0);
272 }
273 #ifndef GRAPHICS_DISABLED
274 if (win != nullptr) {
275 blob->plot(win, ScrollView::RED, ScrollView::RED);
276 }
277 #endif // !GRAPHICS_DISABLED
278 // It is safe to delete the cblob now, as it isn't used by the grid
279 // or BlobOverlapsTooMuch, and the BLOBNBOXes will go away with the
280 // dead_blobs list.
281 // TODO: remove next line, currently still needed for resultiterator_test.
282 delete blob->remove_cblob();
283 dead_it.add_to_end(blob_it.extract());
284 }
285 }
286 }
287
288 // Returns true if the given blob overlaps more than max_overlaps blobs
289 // in the current grid.
290 bool CCNonTextDetect::BlobOverlapsTooMuch(BLOBNBOX *blob, int max_overlaps) {
291 // Search the grid to see what intersects it.
292 // Setup a Rectangle search for overlapping this blob.
293 BlobGridSearch rsearch(this);
294 const TBOX &box = blob->bounding_box();
295 rsearch.StartRectSearch(box);
296 rsearch.SetUniqueMode(true);
297 BLOBNBOX *neighbour;
298 int overlap_count = 0;
299 while (overlap_count <= max_overlaps && (neighbour = rsearch.NextRectSearch()) != nullptr) {
300 if (box.major_overlap(neighbour->bounding_box())) {
301 ++overlap_count;
302 if (overlap_count > max_overlaps) {
303 return true;
304 }
305 }
306 }
307 return false;
308 }
309
310 } // namespace tesseract.