2 * Copyright (c) 2005 Robert Edele <yartrebo@earthlink.net>
3 * Copyright (c) 2012 Stefano Sabatini
5 * This file is part of FFmpeg.
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.
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.
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
24 * Advanced blur-based logo removing filter
26 * This filter loads an image mask file showing where a logo is and
27 * uses a blur transform to remove the logo.
29 * Based on the libmpcodecs remove-logo filter by Robert Edele.
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.
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.
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.
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.
72 #include "libavutil/imgutils.h"
77 #include "lavfutils.h"
78 #include "lswsutils.h"
81 /* Stores our collection of masks. The first is for an array of
82 the second for the y axis, and the third for the x axis. */
87 uint8_t *full_mask_data;
88 FFBoundingBox full_mask_bbox;
89 uint8_t *half_mask_data;
90 FFBoundingBox half_mask_bbox;
94 * Choose a slightly larger mask size to improve performance.
96 * This function maps the absolute minimum mask size needed to the
97 * mask size we'll actually use. f(x) = x (the smallest that will
98 * work) will produce the sharpest results, but will be quite
99 * jittery. f(x) = 1.25x (what I'm using) is a good tradeoff in my
100 * opinion. This will calculate only at init-time, so you can put a
101 * long expression here without effecting performance.
103 #define apply_mask_fudge_factor(x) (((x) >> 2) + x)
106 * Pre-process an image to give distance information.
108 * This function takes a bitmap image and converts it in place into a
109 * distance image. A distance image is zero for pixels outside of the
110 * logo and is the Manhattan distance (|dx| + |dy|) from the logo edge
111 * for pixels inside of the logo. This will overestimate the distance,
112 * but that is safe, and is far easier to implement than a proper
113 * pythagorean distance since I'm using a modified erosion algorithm
114 * to compute the distances.
116 * @param mask image which will be converted from a greyscale image
117 * into a distance image.
119 static void convert_mask_to_strength_mask(uint8_t *data, int linesize,
120 int w, int h, int min_val,
125 /* How many times we've gone through the loop. Used in the
126 in-place erosion algorithm and to get us max_mask_size later on. */
127 int current_pass = 0;
129 /* set all non-zero values to 1 */
130 for (y = 0; y < h; y++)
131 for (x = 0; x < w; x++)
132 data[y*linesize + x] = data[y*linesize + x] > min_val;
134 /* For each pass, if a pixel is itself the same value as the
135 current pass, and its four neighbors are too, then it is
136 incremented. If no pixels are incremented by the end of the
137 pass, then we go again. Edge pixels are counted as always
138 excluded (this should be true anyway for any sane mask, but if
139 it isn't this will ensure that we eventually exit). */
141 /* If this doesn't get set by the end of this pass, then we're done. */
142 int has_anything_changed = 0;
143 uint8_t *current_pixel0 = data, *current_pixel;
146 for (y = 1; y < h-1; y++) {
147 current_pixel = current_pixel0;
148 for (x = 1; x < w-1; x++) {
149 /* Apply the in-place erosion transform. It is based
150 on the following two premises:
151 1 - Any pixel that fails 1 erosion will fail all
154 2 - Only pixels having survived all erosions up to
155 the present will be >= to current_pass.
156 It doesn't matter if it survived the current pass,
157 failed it, or hasn't been tested yet. By using >=
158 instead of ==, we allow the algorithm to work in
160 if ( *current_pixel >= current_pass &&
161 *(current_pixel + 1) >= current_pass &&
162 *(current_pixel - 1) >= current_pass &&
163 *(current_pixel + w) >= current_pass &&
164 *(current_pixel - w) >= current_pass) {
165 /* Increment the value since it still has not been
166 * eroded, as evidenced by the if statement that
167 * just evaluated to true. */
169 has_anything_changed = 1;
173 current_pixel0 += linesize;
175 if (!has_anything_changed)
179 /* Apply the fudge factor, which will increase the size of the
180 * mask a little to reduce jitter at the cost of more blur. */
181 for (y = 1; y < h - 1; y++)
182 for (x = 1; x < w - 1; x++)
183 data[(y * linesize) + x] = apply_mask_fudge_factor(data[(y * linesize) + x]);
185 /* As a side-effect, we now know the maximum mask size, which
186 * we'll use to generate our masks. */
187 /* Apply the fudge factor to this number too, since we must ensure
188 * that enough masks are generated. */
189 *max_mask_size = apply_mask_fudge_factor(current_pass + 1);
192 static int query_formats(AVFilterContext *ctx)
194 enum AVPixelFormat pix_fmts[] = { AV_PIX_FMT_YUV420P, AV_PIX_FMT_NONE };
195 ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
199 static int load_mask(uint8_t **mask, int *w, int *h,
200 const char *filename, void *log_ctx)
203 enum AVPixelFormat pix_fmt;
204 uint8_t *src_data[4], *gray_data[4];
205 int src_linesize[4], gray_linesize[4];
207 /* load image from file */
208 if ((ret = ff_load_image(src_data, src_linesize, w, h, &pix_fmt, filename, log_ctx)) < 0)
211 /* convert the image to GRAY8 */
212 if ((ret = ff_scale_image(gray_data, gray_linesize, *w, *h, AV_PIX_FMT_GRAY8,
213 src_data, src_linesize, *w, *h, pix_fmt,
217 /* copy mask to a newly allocated array */
218 *mask = av_malloc(*w * *h);
220 ret = AVERROR(ENOMEM);
221 av_image_copy_plane(*mask, *w, gray_data[0], gray_linesize[0], *w, *h);
224 av_free(src_data[0]);
225 av_free(gray_data[0]);
230 * Generate a scaled down image with half width, height, and intensity.
232 * This function not only scales down an image, but halves the value
233 * in each pixel too. The purpose of this is to produce a chroma
234 * filter image out of a luma filter image. The pixel values store the
235 * distance to the edge of the logo and halving the dimensions halves
236 * the distance. This function rounds up, because a downwards rounding
237 * error could cause the filter to fail, but an upwards rounding error
238 * will only cause a minor amount of excess blur in the chroma planes.
240 static void generate_half_size_image(const uint8_t *src_data, int src_linesize,
241 uint8_t *dst_data, int dst_linesize,
242 int src_w, int src_h,
247 /* Copy over the image data, using the average of 4 pixels for to
248 * calculate each downsampled pixel. */
249 for (y = 0; y < src_h/2; y++) {
250 for (x = 0; x < src_w/2; x++) {
251 /* Set the pixel if there exists a non-zero value in the
252 * source pixels, else clear it. */
253 dst_data[(y * dst_linesize) + x] =
254 src_data[((y << 1) * src_linesize) + (x << 1)] ||
255 src_data[((y << 1) * src_linesize) + (x << 1) + 1] ||
256 src_data[(((y << 1) + 1) * src_linesize) + (x << 1)] ||
257 src_data[(((y << 1) + 1) * src_linesize) + (x << 1) + 1];
258 dst_data[(y * dst_linesize) + x] = FFMIN(1, dst_data[(y * dst_linesize) + x]);
262 convert_mask_to_strength_mask(dst_data, dst_linesize,
263 src_w/2, src_h/2, 0, max_mask_size);
266 static av_cold int init(AVFilterContext *ctx, const char *args)
268 RemovelogoContext *removelogo = ctx->priv;
272 int full_max_mask_size, half_max_mask_size;
275 av_log(ctx, AV_LOG_ERROR, "An image file must be specified as argument\n");
276 return AVERROR(EINVAL);
279 /* Load our mask image. */
280 if ((ret = load_mask(&removelogo->full_mask_data, &w, &h, args, ctx)) < 0)
282 removelogo->mask_w = w;
283 removelogo->mask_h = h;
285 convert_mask_to_strength_mask(removelogo->full_mask_data, w, w, h,
286 16, &full_max_mask_size);
288 /* Create the scaled down mask image for the chroma planes. */
289 if (!(removelogo->half_mask_data = av_mallocz(w/2 * h/2)))
290 return AVERROR(ENOMEM);
291 generate_half_size_image(removelogo->full_mask_data, w,
292 removelogo->half_mask_data, w/2,
293 w, h, &half_max_mask_size);
295 removelogo->max_mask_size = FFMAX(full_max_mask_size, half_max_mask_size);
297 /* Create a circular mask for each size up to max_mask_size. When
298 the filter is applied, the mask size is determined on a pixel
299 by pixel basis, with pixels nearer the edge of the logo getting
300 smaller mask sizes. */
301 mask = (int ***)av_malloc(sizeof(int **) * (removelogo->max_mask_size + 1));
303 return AVERROR(ENOMEM);
305 for (a = 0; a <= removelogo->max_mask_size; a++) {
306 mask[a] = (int **)av_malloc(sizeof(int *) * ((a * 2) + 1));
308 return AVERROR(ENOMEM);
309 for (b = -a; b <= a; b++) {
310 mask[a][b + a] = (int *)av_malloc(sizeof(int) * ((a * 2) + 1));
312 return AVERROR(ENOMEM);
313 for (c = -a; c <= a; c++) {
314 if ((b * b) + (c * c) <= (a * a)) /* Circular 0/1 mask. */
315 mask[a][b + a][c + a] = 1;
317 mask[a][b + a][c + a] = 0;
321 removelogo->mask = mask;
323 /* Calculate our bounding rectangles, which determine in what
324 * region the logo resides for faster processing. */
325 ff_calculate_bounding_box(&removelogo->full_mask_bbox, removelogo->full_mask_data, w, w, h, 0);
326 ff_calculate_bounding_box(&removelogo->half_mask_bbox, removelogo->half_mask_data, w/2, w/2, h/2, 0);
328 #define SHOW_LOGO_INFO(mask_type) \
329 av_log(ctx, AV_LOG_VERBOSE, #mask_type " x1:%d x2:%d y1:%d y2:%d max_mask_size:%d\n", \
330 removelogo->mask_type##_mask_bbox.x1, removelogo->mask_type##_mask_bbox.x2, \
331 removelogo->mask_type##_mask_bbox.y1, removelogo->mask_type##_mask_bbox.y2, \
332 mask_type##_max_mask_size);
333 SHOW_LOGO_INFO(full);
334 SHOW_LOGO_INFO(half);
339 static int config_props_input(AVFilterLink *inlink)
341 AVFilterContext *ctx = inlink->dst;
342 RemovelogoContext *removelogo = ctx->priv;
344 if (inlink->w != removelogo->mask_w || inlink->h != removelogo->mask_h) {
345 av_log(ctx, AV_LOG_INFO,
346 "Mask image size %dx%d does not match with the input video size %dx%d\n",
347 removelogo->mask_w, removelogo->mask_h, inlink->w, inlink->h);
348 return AVERROR(EINVAL);
357 * It takes a pixel that is inside the mask and blurs it. It does so
358 * by finding the average of all the pixels within the mask and
359 * outside of the mask.
361 * @param mask_data the mask plane to use for averaging
362 * @param image_data the image plane to blur
363 * @param w width of the image
364 * @param h height of the image
365 * @param x x-coordinate of the pixel to blur
366 * @param y y-coordinate of the pixel to blur
368 static unsigned int blur_pixel(int ***mask,
369 const uint8_t *mask_data, int mask_linesize,
370 uint8_t *image_data, int image_linesize,
371 int w, int h, int x, int y)
373 /* Mask size tells how large a circle to use. The radius is about
374 * (slightly larger than) mask size. */
376 int start_posx, start_posy, end_posx, end_posy;
378 unsigned int accumulator = 0, divisor = 0;
379 /* What pixel we are reading out of the circular blur mask. */
380 const uint8_t *image_read_position;
381 /* What pixel we are reading out of the filter image. */
382 const uint8_t *mask_read_position;
384 /* Prepare our bounding rectangle and clip it if need be. */
385 mask_size = mask_data[y * mask_linesize + x];
386 start_posx = FFMAX(0, x - mask_size);
387 start_posy = FFMAX(0, y - mask_size);
388 end_posx = FFMIN(w - 1, x + mask_size);
389 end_posy = FFMIN(h - 1, y + mask_size);
391 image_read_position = image_data + image_linesize * start_posy + start_posx;
392 mask_read_position = mask_data + mask_linesize * start_posy + start_posx;
394 for (j = start_posy; j <= end_posy; j++) {
395 for (i = start_posx; i <= end_posx; i++) {
396 /* Check if this pixel is in the mask or not. Only use the
397 * pixel if it is not. */
398 if (!(*mask_read_position) && mask[mask_size][i - start_posx][j - start_posy]) {
399 accumulator += *image_read_position;
403 image_read_position++;
404 mask_read_position++;
407 image_read_position += (image_linesize - ((end_posx + 1) - start_posx));
408 mask_read_position += (mask_linesize - ((end_posx + 1) - start_posx));
411 /* If divisor is 0, it means that not a single pixel is outside of
412 the logo, so we have no data. Else we need to normalise the
413 data using the divisor. */
414 return divisor == 0 ? 255:
415 (accumulator + (divisor / 2)) / divisor; /* divide, taking into account average rounding error */
419 * Blur image plane using a mask.
421 * @param source The image to have it's logo removed.
422 * @param destination Where the output image will be stored.
423 * @param source_stride How far apart (in memory) two consecutive lines are.
424 * @param destination Same as source_stride, but for the destination image.
425 * @param width Width of the image. This is the same for source and destination.
426 * @param height Height of the image. This is the same for source and destination.
427 * @param is_image_direct If the image is direct, then source and destination are
428 * the same and we can save a lot of time by not copying pixels that
430 * @param filter The image that stores the distance to the edge of the logo for
432 * @param logo_start_x smallest x-coordinate that contains at least 1 logo pixel.
433 * @param logo_start_y smallest y-coordinate that contains at least 1 logo pixel.
434 * @param logo_end_x largest x-coordinate that contains at least 1 logo pixel.
435 * @param logo_end_y largest y-coordinate that contains at least 1 logo pixel.
437 * This function processes an entire plane. Pixels outside of the logo are copied
438 * to the output without change, and pixels inside the logo have the de-blurring
441 static void blur_image(int ***mask,
442 const uint8_t *src_data, int src_linesize,
443 uint8_t *dst_data, int dst_linesize,
444 const uint8_t *mask_data, int mask_linesize,
445 int w, int h, int direct,
450 const uint8_t *src_line;
453 av_image_copy_plane(dst_data, dst_linesize, src_data, src_linesize, w, h);
455 for (y = bbox->y1; y <= bbox->y2; y++) {
456 src_line = src_data + src_linesize * y;
457 dst_line = dst_data + dst_linesize * y;
459 for (x = bbox->x1; x <= bbox->x2; x++) {
460 if (mask_data[y * mask_linesize + x]) {
461 /* Only process if we are in the mask. */
462 dst_line[x] = blur_pixel(mask,
463 mask_data, mask_linesize,
464 dst_data, dst_linesize,
467 /* Else just copy the data. */
469 dst_line[x] = src_line[x];
475 static int start_frame(AVFilterLink *inlink, AVFilterBufferRef *inpicref)
477 AVFilterLink *outlink = inlink->dst->outputs[0];
478 AVFilterBufferRef *outpicref;
480 outpicref = inpicref;
482 outlink->out_buf = outpicref;
483 return ff_start_frame(outlink, avfilter_ref_buffer(outpicref, ~0));
486 static int end_frame(AVFilterLink *inlink)
488 RemovelogoContext *removelogo = inlink->dst->priv;
489 AVFilterLink *outlink = inlink->dst->outputs[0];
490 AVFilterBufferRef *inpicref = inlink ->cur_buf;
491 AVFilterBufferRef *outpicref = outlink->out_buf;
492 int direct = inpicref == outpicref;
494 blur_image(removelogo->mask,
495 inpicref ->data[0], inpicref ->linesize[0],
496 outpicref->data[0], outpicref->linesize[0],
497 removelogo->full_mask_data, inlink->w,
498 inlink->w, inlink->h, direct, &removelogo->full_mask_bbox);
499 blur_image(removelogo->mask,
500 inpicref ->data[1], inpicref ->linesize[1],
501 outpicref->data[1], outpicref->linesize[1],
502 removelogo->half_mask_data, inlink->w/2,
503 inlink->w/2, inlink->h/2, direct, &removelogo->half_mask_bbox);
504 blur_image(removelogo->mask,
505 inpicref ->data[2], inpicref ->linesize[2],
506 outpicref->data[2], outpicref->linesize[2],
507 removelogo->half_mask_data, inlink->w/2,
508 inlink->w/2, inlink->h/2, direct, &removelogo->half_mask_bbox);
510 ff_draw_slice(outlink, 0, inlink->h, 1);
511 return ff_end_frame(outlink);
514 static void uninit(AVFilterContext *ctx)
516 RemovelogoContext *removelogo = ctx->priv;
519 av_freep(&removelogo->full_mask_data);
520 av_freep(&removelogo->half_mask_data);
522 if (removelogo->mask) {
523 /* Loop through each mask. */
524 for (a = 0; a <= removelogo->max_mask_size; a++) {
525 /* Loop through each scanline in a mask. */
526 for (b = -a; b <= a; b++) {
527 av_free(removelogo->mask[a][b + a]); /* Free a scanline. */
529 av_free(removelogo->mask[a]);
531 /* Free the array of pointers pointing to the masks. */
532 av_freep(&removelogo->mask);
536 static int null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir) { return 0; }
538 AVFilter avfilter_vf_removelogo = {
539 .name = "removelogo",
540 .description = NULL_IF_CONFIG_SMALL("Remove a TV logo based on a mask image."),
541 .priv_size = sizeof(RemovelogoContext),
544 .query_formats = query_formats,
546 .inputs = (const AVFilterPad[]) {
548 .type = AVMEDIA_TYPE_VIDEO,
549 .get_video_buffer = ff_null_get_video_buffer,
550 .config_props = config_props_input,
551 .draw_slice = null_draw_slice,
552 .start_frame = start_frame,
553 .end_frame = end_frame,
554 .min_perms = AV_PERM_WRITE | AV_PERM_READ },
557 .outputs = (const AVFilterPad[]) {
559 .type = AVMEDIA_TYPE_VIDEO, },