summaryrefslogtreecommitdiff
path: root/libavfilter/vf_removelogo.c (plain)
blob: 94b92a5853409523d36a78d347eaae4fd019fee9
1/*
2 * Copyright (c) 2005 Robert Edele <yartrebo@earthlink.net>
3 * Copyright (c) 2012 Stefano Sabatini
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22/**
23 * @file
24 * Advanced blur-based logo removing filter
25 *
26 * This filter loads an image mask file showing where a logo is and
27 * uses a blur transform to remove the logo.
28 *
29 * Based on the libmpcodecs remove-logo filter by Robert Edele.
30 */
31
32/**
33 * This code implements a filter to remove annoying TV logos and other annoying
34 * images placed onto a video stream. It works by filling in the pixels that
35 * comprise the logo with neighboring pixels. The transform is very loosely
36 * based on a gaussian blur, but it is different enough to merit its own
37 * paragraph later on. It is a major improvement on the old delogo filter as it
38 * both uses a better blurring algorithm and uses a bitmap to use an arbitrary
39 * and generally much tighter fitting shape than a rectangle.
40 *
41 * The logo removal algorithm has two key points. The first is that it
42 * distinguishes between pixels in the logo and those not in the logo by using
43 * the passed-in bitmap. Pixels not in the logo are copied over directly without
44 * being modified and they also serve as source pixels for the logo
45 * fill-in. Pixels inside the logo have the mask applied.
46 *
47 * At init-time the bitmap is reprocessed internally, and the distance to the
48 * nearest edge of the logo (Manhattan distance), along with a little extra to
49 * remove rough edges, is stored in each pixel. This is done using an in-place
50 * erosion algorithm, and incrementing each pixel that survives any given
51 * erosion. Once every pixel is eroded, the maximum value is recorded, and a
52 * set of masks from size 0 to this size are generaged. The masks are circular
53 * binary masks, where each pixel within a radius N (where N is the size of the
54 * mask) is a 1, and all other pixels are a 0. Although a gaussian mask would be
55 * more mathematically accurate, a binary mask works better in practice because
56 * we generally do not use the central pixels in the mask (because they are in
57 * the logo region), and thus a gaussian mask will cause too little blur and
58 * thus a very unstable image.
59 *
60 * The mask is applied in a special way. Namely, only pixels in the mask that
61 * line up to pixels outside the logo are used. The dynamic mask size means that
62 * the mask is just big enough so that the edges touch pixels outside the logo,
63 * so the blurring is kept to a minimum and at least the first boundary
64 * condition is met (that the image function itself is continuous), even if the
65 * second boundary condition (that the derivative of the image function is
66 * continuous) is not met. A masking algorithm that does preserve the second
67 * boundary coundition (perhaps something based on a highly-modified bi-cubic
68 * algorithm) should offer even better results on paper, but the noise in a
69 * typical TV signal should make anything based on derivatives hopelessly noisy.
70 */
71
72#include "libavutil/imgutils.h"
73#include "libavutil/opt.h"
74#include "avfilter.h"
75#include "formats.h"
76#include "internal.h"
77#include "video.h"
78#include "bbox.h"
79#include "lavfutils.h"
80#include "lswsutils.h"
81
82typedef struct {
83 const AVClass *class;
84 char *filename;
85 /* Stores our collection of masks. The first is for an array of
86 the second for the y axis, and the third for the x axis. */
87 int ***mask;
88 int max_mask_size;
89 int mask_w, mask_h;
90
91 uint8_t *full_mask_data;
92 FFBoundingBox full_mask_bbox;
93 uint8_t *half_mask_data;
94 FFBoundingBox half_mask_bbox;
95} RemovelogoContext;
96
97#define OFFSET(x) offsetof(RemovelogoContext, x)
98#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
99static const AVOption removelogo_options[] = {
100 { "filename", "set bitmap filename", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
101 { "f", "set bitmap filename", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
102 { NULL }
103};
104
105AVFILTER_DEFINE_CLASS(removelogo);
106
107/**
108 * Choose a slightly larger mask size to improve performance.
109 *
110 * This function maps the absolute minimum mask size needed to the
111 * mask size we'll actually use. f(x) = x (the smallest that will
112 * work) will produce the sharpest results, but will be quite
113 * jittery. f(x) = 1.25x (what I'm using) is a good tradeoff in my
114 * opinion. This will calculate only at init-time, so you can put a
115 * long expression here without effecting performance.
116 */
117#define apply_mask_fudge_factor(x) (((x) >> 2) + (x))
118
119/**
120 * Pre-process an image to give distance information.
121 *
122 * This function takes a bitmap image and converts it in place into a
123 * distance image. A distance image is zero for pixels outside of the
124 * logo and is the Manhattan distance (|dx| + |dy|) from the logo edge
125 * for pixels inside of the logo. This will overestimate the distance,
126 * but that is safe, and is far easier to implement than a proper
127 * pythagorean distance since I'm using a modified erosion algorithm
128 * to compute the distances.
129 *
130 * @param mask image which will be converted from a greyscale image
131 * into a distance image.
132 */
133static void convert_mask_to_strength_mask(uint8_t *data, int linesize,
134 int w, int h, int min_val,
135 int *max_mask_size)
136{
137 int x, y;
138
139 /* How many times we've gone through the loop. Used in the
140 in-place erosion algorithm and to get us max_mask_size later on. */
141 int current_pass = 0;
142
143 /* set all non-zero values to 1 */
144 for (y = 0; y < h; y++)
145 for (x = 0; x < w; x++)
146 data[y*linesize + x] = data[y*linesize + x] > min_val;
147
148 /* For each pass, if a pixel is itself the same value as the
149 current pass, and its four neighbors are too, then it is
150 incremented. If no pixels are incremented by the end of the
151 pass, then we go again. Edge pixels are counted as always
152 excluded (this should be true anyway for any sane mask, but if
153 it isn't this will ensure that we eventually exit). */
154 while (1) {
155 /* If this doesn't get set by the end of this pass, then we're done. */
156 int has_anything_changed = 0;
157 uint8_t *current_pixel0 = data + 1 + linesize, *current_pixel;
158 current_pass++;
159
160 for (y = 1; y < h-1; y++) {
161 current_pixel = current_pixel0;
162 for (x = 1; x < w-1; x++) {
163 /* Apply the in-place erosion transform. It is based
164 on the following two premises:
165 1 - Any pixel that fails 1 erosion will fail all
166 future erosions.
167
168 2 - Only pixels having survived all erosions up to
169 the present will be >= to current_pass.
170 It doesn't matter if it survived the current pass,
171 failed it, or hasn't been tested yet. By using >=
172 instead of ==, we allow the algorithm to work in
173 place. */
174 if ( *current_pixel >= current_pass &&
175 *(current_pixel + 1) >= current_pass &&
176 *(current_pixel - 1) >= current_pass &&
177 *(current_pixel + linesize) >= current_pass &&
178 *(current_pixel - linesize) >= current_pass) {
179 /* Increment the value since it still has not been
180 * eroded, as evidenced by the if statement that
181 * just evaluated to true. */
182 (*current_pixel)++;
183 has_anything_changed = 1;
184 }
185 current_pixel++;
186 }
187 current_pixel0 += linesize;
188 }
189 if (!has_anything_changed)
190 break;
191 }
192
193 /* Apply the fudge factor, which will increase the size of the
194 * mask a little to reduce jitter at the cost of more blur. */
195 for (y = 1; y < h - 1; y++)
196 for (x = 1; x < w - 1; x++)
197 data[(y * linesize) + x] = apply_mask_fudge_factor(data[(y * linesize) + x]);
198
199 /* As a side-effect, we now know the maximum mask size, which
200 * we'll use to generate our masks. */
201 /* Apply the fudge factor to this number too, since we must ensure
202 * that enough masks are generated. */
203 *max_mask_size = apply_mask_fudge_factor(current_pass + 1);
204}
205
206static int query_formats(AVFilterContext *ctx)
207{
208 static const enum AVPixelFormat pix_fmts[] = { AV_PIX_FMT_YUV420P, AV_PIX_FMT_NONE };
209 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
210 if (!fmts_list)
211 return AVERROR(ENOMEM);
212 return ff_set_common_formats(ctx, fmts_list);
213}
214
215static int load_mask(uint8_t **mask, int *w, int *h,
216 const char *filename, void *log_ctx)
217{
218 int ret;
219 enum AVPixelFormat pix_fmt;
220 uint8_t *src_data[4], *gray_data[4];
221 int src_linesize[4], gray_linesize[4];
222
223 /* load image from file */
224 if ((ret = ff_load_image(src_data, src_linesize, w, h, &pix_fmt, filename, log_ctx)) < 0)
225 return ret;
226
227 /* convert the image to GRAY8 */
228 if ((ret = ff_scale_image(gray_data, gray_linesize, *w, *h, AV_PIX_FMT_GRAY8,
229 src_data, src_linesize, *w, *h, pix_fmt,
230 log_ctx)) < 0)
231 goto end;
232
233 /* copy mask to a newly allocated array */
234 *mask = av_malloc(*w * *h);
235 if (!*mask)
236 ret = AVERROR(ENOMEM);
237 av_image_copy_plane(*mask, *w, gray_data[0], gray_linesize[0], *w, *h);
238
239end:
240 av_freep(&src_data[0]);
241 av_freep(&gray_data[0]);
242 return ret;
243}
244
245/**
246 * Generate a scaled down image with half width, height, and intensity.
247 *
248 * This function not only scales down an image, but halves the value
249 * in each pixel too. The purpose of this is to produce a chroma
250 * filter image out of a luma filter image. The pixel values store the
251 * distance to the edge of the logo and halving the dimensions halves
252 * the distance. This function rounds up, because a downwards rounding
253 * error could cause the filter to fail, but an upwards rounding error
254 * will only cause a minor amount of excess blur in the chroma planes.
255 */
256static void generate_half_size_image(const uint8_t *src_data, int src_linesize,
257 uint8_t *dst_data, int dst_linesize,
258 int src_w, int src_h,
259 int *max_mask_size)
260{
261 int x, y;
262
263 /* Copy over the image data, using the average of 4 pixels for to
264 * calculate each downsampled pixel. */
265 for (y = 0; y < src_h/2; y++) {
266 for (x = 0; x < src_w/2; x++) {
267 /* Set the pixel if there exists a non-zero value in the
268 * source pixels, else clear it. */
269 dst_data[(y * dst_linesize) + x] =
270 src_data[((y << 1) * src_linesize) + (x << 1)] ||
271 src_data[((y << 1) * src_linesize) + (x << 1) + 1] ||
272 src_data[(((y << 1) + 1) * src_linesize) + (x << 1)] ||
273 src_data[(((y << 1) + 1) * src_linesize) + (x << 1) + 1];
274 dst_data[(y * dst_linesize) + x] = FFMIN(1, dst_data[(y * dst_linesize) + x]);
275 }
276 }
277
278 convert_mask_to_strength_mask(dst_data, dst_linesize,
279 src_w/2, src_h/2, 0, max_mask_size);
280}
281
282static av_cold int init(AVFilterContext *ctx)
283{
284 RemovelogoContext *s = ctx->priv;
285 int ***mask;
286 int ret = 0;
287 int a, b, c, w, h;
288 int full_max_mask_size, half_max_mask_size;
289
290 if (!s->filename) {
291 av_log(ctx, AV_LOG_ERROR, "The bitmap file name is mandatory\n");
292 return AVERROR(EINVAL);
293 }
294
295 /* Load our mask image. */
296 if ((ret = load_mask(&s->full_mask_data, &w, &h, s->filename, ctx)) < 0)
297 return ret;
298 s->mask_w = w;
299 s->mask_h = h;
300
301 convert_mask_to_strength_mask(s->full_mask_data, w, w, h,
302 16, &full_max_mask_size);
303
304 /* Create the scaled down mask image for the chroma planes. */
305 if (!(s->half_mask_data = av_mallocz(w/2 * h/2)))
306 return AVERROR(ENOMEM);
307 generate_half_size_image(s->full_mask_data, w,
308 s->half_mask_data, w/2,
309 w, h, &half_max_mask_size);
310
311 s->max_mask_size = FFMAX(full_max_mask_size, half_max_mask_size);
312
313 /* Create a circular mask for each size up to max_mask_size. When
314 the filter is applied, the mask size is determined on a pixel
315 by pixel basis, with pixels nearer the edge of the logo getting
316 smaller mask sizes. */
317 mask = (int ***)av_malloc_array(s->max_mask_size + 1, sizeof(int **));
318 if (!mask)
319 return AVERROR(ENOMEM);
320
321 for (a = 0; a <= s->max_mask_size; a++) {
322 mask[a] = (int **)av_malloc_array((a * 2) + 1, sizeof(int *));
323 if (!mask[a]) {
324 av_free(mask);
325 return AVERROR(ENOMEM);
326 }
327 for (b = -a; b <= a; b++) {
328 mask[a][b + a] = (int *)av_malloc_array((a * 2) + 1, sizeof(int));
329 if (!mask[a][b + a]) {
330 av_free(mask);
331 return AVERROR(ENOMEM);
332 }
333 for (c = -a; c <= a; c++) {
334 if ((b * b) + (c * c) <= (a * a)) /* Circular 0/1 mask. */
335 mask[a][b + a][c + a] = 1;
336 else
337 mask[a][b + a][c + a] = 0;
338 }
339 }
340 }
341 s->mask = mask;
342
343 /* Calculate our bounding rectangles, which determine in what
344 * region the logo resides for faster processing. */
345 ff_calculate_bounding_box(&s->full_mask_bbox, s->full_mask_data, w, w, h, 0);
346 ff_calculate_bounding_box(&s->half_mask_bbox, s->half_mask_data, w/2, w/2, h/2, 0);
347
348#define SHOW_LOGO_INFO(mask_type) \
349 av_log(ctx, AV_LOG_VERBOSE, #mask_type " x1:%d x2:%d y1:%d y2:%d max_mask_size:%d\n", \
350 s->mask_type##_mask_bbox.x1, s->mask_type##_mask_bbox.x2, \
351 s->mask_type##_mask_bbox.y1, s->mask_type##_mask_bbox.y2, \
352 mask_type##_max_mask_size);
353 SHOW_LOGO_INFO(full);
354 SHOW_LOGO_INFO(half);
355
356 return 0;
357}
358
359static int config_props_input(AVFilterLink *inlink)
360{
361 AVFilterContext *ctx = inlink->dst;
362 RemovelogoContext *s = ctx->priv;
363
364 if (inlink->w != s->mask_w || inlink->h != s->mask_h) {
365 av_log(ctx, AV_LOG_INFO,
366 "Mask image size %dx%d does not match with the input video size %dx%d\n",
367 s->mask_w, s->mask_h, inlink->w, inlink->h);
368 return AVERROR(EINVAL);
369 }
370
371 return 0;
372}
373
374/**
375 * Blur image.
376 *
377 * It takes a pixel that is inside the mask and blurs it. It does so
378 * by finding the average of all the pixels within the mask and
379 * outside of the mask.
380 *
381 * @param mask_data the mask plane to use for averaging
382 * @param image_data the image plane to blur
383 * @param w width of the image
384 * @param h height of the image
385 * @param x x-coordinate of the pixel to blur
386 * @param y y-coordinate of the pixel to blur
387 */
388static unsigned int blur_pixel(int ***mask,
389 const uint8_t *mask_data, int mask_linesize,
390 uint8_t *image_data, int image_linesize,
391 int w, int h, int x, int y)
392{
393 /* Mask size tells how large a circle to use. The radius is about
394 * (slightly larger than) mask size. */
395 int mask_size;
396 int start_posx, start_posy, end_posx, end_posy;
397 int i, j;
398 unsigned int accumulator = 0, divisor = 0;
399 /* What pixel we are reading out of the circular blur mask. */
400 const uint8_t *image_read_position;
401 /* What pixel we are reading out of the filter image. */
402 const uint8_t *mask_read_position;
403
404 /* Prepare our bounding rectangle and clip it if need be. */
405 mask_size = mask_data[y * mask_linesize + x];
406 start_posx = FFMAX(0, x - mask_size);
407 start_posy = FFMAX(0, y - mask_size);
408 end_posx = FFMIN(w - 1, x + mask_size);
409 end_posy = FFMIN(h - 1, y + mask_size);
410
411 image_read_position = image_data + image_linesize * start_posy + start_posx;
412 mask_read_position = mask_data + mask_linesize * start_posy + start_posx;
413
414 for (j = start_posy; j <= end_posy; j++) {
415 for (i = start_posx; i <= end_posx; i++) {
416 /* Check if this pixel is in the mask or not. Only use the
417 * pixel if it is not. */
418 if (!(*mask_read_position) && mask[mask_size][i - start_posx][j - start_posy]) {
419 accumulator += *image_read_position;
420 divisor++;
421 }
422
423 image_read_position++;
424 mask_read_position++;
425 }
426
427 image_read_position += (image_linesize - ((end_posx + 1) - start_posx));
428 mask_read_position += (mask_linesize - ((end_posx + 1) - start_posx));
429 }
430
431 /* If divisor is 0, it means that not a single pixel is outside of
432 the logo, so we have no data. Else we need to normalise the
433 data using the divisor. */
434 return divisor == 0 ? 255:
435 (accumulator + (divisor / 2)) / divisor; /* divide, taking into account average rounding error */
436}
437
438/**
439 * Blur image plane using a mask.
440 *
441 * @param source The image to have it's logo removed.
442 * @param destination Where the output image will be stored.
443 * @param source_stride How far apart (in memory) two consecutive lines are.
444 * @param destination Same as source_stride, but for the destination image.
445 * @param width Width of the image. This is the same for source and destination.
446 * @param height Height of the image. This is the same for source and destination.
447 * @param is_image_direct If the image is direct, then source and destination are
448 * the same and we can save a lot of time by not copying pixels that
449 * haven't changed.
450 * @param filter The image that stores the distance to the edge of the logo for
451 * each pixel.
452 * @param logo_start_x smallest x-coordinate that contains at least 1 logo pixel.
453 * @param logo_start_y smallest y-coordinate that contains at least 1 logo pixel.
454 * @param logo_end_x largest x-coordinate that contains at least 1 logo pixel.
455 * @param logo_end_y largest y-coordinate that contains at least 1 logo pixel.
456 *
457 * This function processes an entire plane. Pixels outside of the logo are copied
458 * to the output without change, and pixels inside the logo have the de-blurring
459 * function applied.
460 */
461static void blur_image(int ***mask,
462 const uint8_t *src_data, int src_linesize,
463 uint8_t *dst_data, int dst_linesize,
464 const uint8_t *mask_data, int mask_linesize,
465 int w, int h, int direct,
466 FFBoundingBox *bbox)
467{
468 int x, y;
469 uint8_t *dst_line;
470 const uint8_t *src_line;
471
472 if (!direct)
473 av_image_copy_plane(dst_data, dst_linesize, src_data, src_linesize, w, h);
474
475 for (y = bbox->y1; y <= bbox->y2; y++) {
476 src_line = src_data + src_linesize * y;
477 dst_line = dst_data + dst_linesize * y;
478
479 for (x = bbox->x1; x <= bbox->x2; x++) {
480 if (mask_data[y * mask_linesize + x]) {
481 /* Only process if we are in the mask. */
482 dst_line[x] = blur_pixel(mask,
483 mask_data, mask_linesize,
484 dst_data, dst_linesize,
485 w, h, x, y);
486 } else {
487 /* Else just copy the data. */
488 if (!direct)
489 dst_line[x] = src_line[x];
490 }
491 }
492 }
493}
494
495static int filter_frame(AVFilterLink *inlink, AVFrame *inpicref)
496{
497 RemovelogoContext *s = inlink->dst->priv;
498 AVFilterLink *outlink = inlink->dst->outputs[0];
499 AVFrame *outpicref;
500 int direct = 0;
501
502 if (av_frame_is_writable(inpicref)) {
503 direct = 1;
504 outpicref = inpicref;
505 } else {
506 outpicref = ff_get_video_buffer(outlink, outlink->w, outlink->h);
507 if (!outpicref) {
508 av_frame_free(&inpicref);
509 return AVERROR(ENOMEM);
510 }
511 av_frame_copy_props(outpicref, inpicref);
512 }
513
514 blur_image(s->mask,
515 inpicref ->data[0], inpicref ->linesize[0],
516 outpicref->data[0], outpicref->linesize[0],
517 s->full_mask_data, inlink->w,
518 inlink->w, inlink->h, direct, &s->full_mask_bbox);
519 blur_image(s->mask,
520 inpicref ->data[1], inpicref ->linesize[1],
521 outpicref->data[1], outpicref->linesize[1],
522 s->half_mask_data, inlink->w/2,
523 inlink->w/2, inlink->h/2, direct, &s->half_mask_bbox);
524 blur_image(s->mask,
525 inpicref ->data[2], inpicref ->linesize[2],
526 outpicref->data[2], outpicref->linesize[2],
527 s->half_mask_data, inlink->w/2,
528 inlink->w/2, inlink->h/2, direct, &s->half_mask_bbox);
529
530 if (!direct)
531 av_frame_free(&inpicref);
532
533 return ff_filter_frame(outlink, outpicref);
534}
535
536static av_cold void uninit(AVFilterContext *ctx)
537{
538 RemovelogoContext *s = ctx->priv;
539 int a, b;
540
541 av_freep(&s->full_mask_data);
542 av_freep(&s->half_mask_data);
543
544 if (s->mask) {
545 /* Loop through each mask. */
546 for (a = 0; a <= s->max_mask_size; a++) {
547 /* Loop through each scanline in a mask. */
548 for (b = -a; b <= a; b++) {
549 av_freep(&s->mask[a][b + a]); /* Free a scanline. */
550 }
551 av_freep(&s->mask[a]);
552 }
553 /* Free the array of pointers pointing to the masks. */
554 av_freep(&s->mask);
555 }
556}
557
558static const AVFilterPad removelogo_inputs[] = {
559 {
560 .name = "default",
561 .type = AVMEDIA_TYPE_VIDEO,
562 .config_props = config_props_input,
563 .filter_frame = filter_frame,
564 },
565 { NULL }
566};
567
568static const AVFilterPad removelogo_outputs[] = {
569 {
570 .name = "default",
571 .type = AVMEDIA_TYPE_VIDEO,
572 },
573 { NULL }
574};
575
576AVFilter ff_vf_removelogo = {
577 .name = "removelogo",
578 .description = NULL_IF_CONFIG_SMALL("Remove a TV logo based on a mask image."),
579 .priv_size = sizeof(RemovelogoContext),
580 .init = init,
581 .uninit = uninit,
582 .query_formats = query_formats,
583 .inputs = removelogo_inputs,
584 .outputs = removelogo_outputs,
585 .priv_class = &removelogo_class,
586 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,
587};
588
generated by cgit at 2024-05-17 20:21:52 (GMT)