2 * Copyright (c) 2003 LeFunGus, lefungus@altern.org
4 * This file is part of FFmpeg
6 * FFmpeg is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License along
17 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 #include "libavutil/imgutils.h"
24 #include "libavutil/attributes.h"
25 #include "libavutil/common.h"
26 #include "libavutil/pixdesc.h"
27 #include "libavutil/intreadwrite.h"
28 #include "libavutil/opt.h"
35 typedef struct VagueDenoiserContext {
62 void (*thresholding)(float *block, const int width, const int height,
63 const int stride, const float threshold,
65 } VagueDenoiserContext;
67 #define OFFSET(x) offsetof(VagueDenoiserContext, x)
68 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_FILTERING_PARAM
69 static const AVOption vaguedenoiser_options[] = {
70 { "threshold", "set filtering strength", OFFSET(threshold), AV_OPT_TYPE_FLOAT, {.dbl=2.}, 0,DBL_MAX, FLAGS },
71 { "method", "set filtering method", OFFSET(method), AV_OPT_TYPE_INT, {.i64=2 }, 0, 2, FLAGS, "method" },
72 { "hard", "hard thresholding", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, FLAGS, "method" },
73 { "soft", "soft thresholding", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, FLAGS, "method" },
74 { "garrote", "garrote thresholding", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, FLAGS, "method" },
75 { "nsteps", "set number of steps", OFFSET(nsteps), AV_OPT_TYPE_INT, {.i64=6 }, 1, 32, FLAGS },
76 { "percent", "set percent of full denoising", OFFSET(percent),AV_OPT_TYPE_FLOAT, {.dbl=85}, 0,100, FLAGS },
77 { "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15 }, 0, 15, FLAGS },
78 { "type", "set threshold type", OFFSET(type), AV_OPT_TYPE_INT, {.i64=0 }, 0, 1, FLAGS, "type" },
79 { "universal", "universal (VisuShrink)", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, FLAGS, "type" },
80 { "bayes", "bayes (BayesShrink)", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, FLAGS, "type" },
84 AVFILTER_DEFINE_CLASS(vaguedenoiser);
88 static const float analysis_low[9] = {
89 0.037828455506995f, -0.023849465019380f, -0.110624404418423f, 0.377402855612654f,
90 0.852698679009403f, 0.377402855612654f, -0.110624404418423f, -0.023849465019380f, 0.037828455506995f
93 static const float analysis_high[7] = {
94 -0.064538882628938f, 0.040689417609558f, 0.418092273222212f, -0.788485616405664f,
95 0.418092273222212f, 0.040689417609558f, -0.064538882628938f
98 static const float synthesis_low[7] = {
99 -0.064538882628938f, -0.040689417609558f, 0.418092273222212f, 0.788485616405664f,
100 0.418092273222212f, -0.040689417609558f, -0.064538882628938f
103 static const float synthesis_high[9] = {
104 -0.037828455506995f, -0.023849465019380f, 0.110624404418423f, 0.377402855612654f,
105 -0.852698679009403f, 0.377402855612654f, 0.110624404418423f, -0.023849465019380f, -0.037828455506995f
108 static int query_formats(AVFilterContext *ctx)
110 static const enum AVPixelFormat pix_fmts[] = {
111 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
112 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
113 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
114 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
115 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
116 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
117 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
119 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
120 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
121 AV_PIX_FMT_YUV440P10,
122 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
123 AV_PIX_FMT_YUV440P12,
124 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
125 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
126 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
127 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
128 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
129 AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
130 AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
131 AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
132 AV_PIX_FMT_GBRAP, AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
135 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
137 return AVERROR(ENOMEM);
138 return ff_set_common_formats(ctx, fmts_list);
141 static int config_input(AVFilterLink *inlink)
143 VagueDenoiserContext *s = inlink->dst->priv;
144 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
145 int p, i, nsteps_width, nsteps_height, nsteps_max;
147 s->depth = desc->comp[0].depth;
148 s->bpc = (s->depth + 7) / 8;
149 s->nb_planes = desc->nb_components;
151 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
152 s->planeheight[0] = s->planeheight[3] = inlink->h;
153 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
154 s->planewidth[0] = s->planewidth[3] = inlink->w;
156 s->block = av_malloc_array(inlink->w * inlink->h, sizeof(*s->block));
157 s->in = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->in));
158 s->out = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->out));
159 s->tmp = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->tmp));
161 if (!s->block || !s->in || !s->out || !s->tmp)
162 return AVERROR(ENOMEM);
164 s->threshold *= 1 << (s->depth - 8);
165 s->peak = (1 << s->depth) - 1;
167 nsteps_width = ((s->planes & 2 || s->planes & 4) && s->nb_planes > 1) ? s->planewidth[1] : s->planewidth[0];
168 nsteps_height = ((s->planes & 2 || s->planes & 4) && s->nb_planes > 1) ? s->planeheight[1] : s->planeheight[0];
170 for (nsteps_max = 1; nsteps_max < 15; nsteps_max++) {
171 if (pow(2, nsteps_max) >= nsteps_width || pow(2, nsteps_max) >= nsteps_height)
175 s->nsteps = FFMIN(s->nsteps, nsteps_max - 2);
177 for (p = 0; p < 4; p++) {
178 s->hlowsize[p][0] = (s->planewidth[p] + 1) >> 1;
179 s->hhighsize[p][0] = s->planewidth[p] >> 1;
180 s->vlowsize[p][0] = (s->planeheight[p] + 1) >> 1;
181 s->vhighsize[p][0] = s->planeheight[p] >> 1;
183 for (i = 1; i < s->nsteps; i++) {
184 s->hlowsize[p][i] = (s->hlowsize[p][i - 1] + 1) >> 1;
185 s->hhighsize[p][i] = s->hlowsize[p][i - 1] >> 1;
186 s->vlowsize[p][i] = (s->vlowsize[p][i - 1] + 1) >> 1;
187 s->vhighsize[p][i] = s->vlowsize[p][i - 1] >> 1;
194 static inline void copy(const float *p1, float *p2, const int length)
196 memcpy(p2, p1, length * sizeof(float));
199 static inline void copyv(const float *p1, const int stride1, float *p2, const int length)
203 for (i = 0; i < length; i++) {
209 static inline void copyh(const float *p1, float *p2, const int stride2, const int length)
213 for (i = 0; i < length; i++) {
219 // Do symmetric extension of data using prescribed symmetries
220 // Original values are in output[npad] through output[npad+size-1]
221 // New values will be placed in output[0] through output[npad] and in output[npad+size] through output[2*npad+size-1] (note: end values may not be filled in)
222 // extension at left bdry is ... 3 2 1 0 | 0 1 2 3 ...
223 // same for right boundary
224 // if right_ext=1 then ... 3 2 1 0 | 1 2 3
225 static void symmetric_extension(float *output, const int size, const int left_ext, const int right_ext)
228 int last = NPAD - 1 + size;
229 const int originalLast = last;
233 output[--first] = output[NPAD];
235 output[++last] = output[originalLast];
239 for (i = 0; i < nextend; i++)
240 output[--first] = output[NPAD + 1 + i];
242 idx = NPAD + NPAD - 1 + size;
245 nextend = idx - last;
246 for (i = 0; i < nextend; i++)
247 output[++last] = output[originalLast - 1 - i];
250 static void transform_step(float *input, float *output, const int size, const int low_size, VagueDenoiserContext *s)
254 symmetric_extension(input, size, 1, 1);
256 for (i = NPAD; i < NPAD + low_size; i++) {
257 const float a = input[2 * i - 14] * analysis_low[0];
258 const float b = input[2 * i - 13] * analysis_low[1];
259 const float c = input[2 * i - 12] * analysis_low[2];
260 const float d = input[2 * i - 11] * analysis_low[3];
261 const float e = input[2 * i - 10] * analysis_low[4];
262 const float f = input[2 * i - 9] * analysis_low[3];
263 const float g = input[2 * i - 8] * analysis_low[2];
264 const float h = input[2 * i - 7] * analysis_low[1];
265 const float k = input[2 * i - 6] * analysis_low[0];
267 output[i] = a + b + c + d + e + f + g + h + k;
270 for (i = NPAD; i < NPAD + low_size; i++) {
271 const float a = input[2 * i - 12] * analysis_high[0];
272 const float b = input[2 * i - 11] * analysis_high[1];
273 const float c = input[2 * i - 10] * analysis_high[2];
274 const float d = input[2 * i - 9] * analysis_high[3];
275 const float e = input[2 * i - 8] * analysis_high[2];
276 const float f = input[2 * i - 7] * analysis_high[1];
277 const float g = input[2 * i - 6] * analysis_high[0];
279 output[i + low_size] = a + b + c + d + e + f + g;
283 static void invert_step(const float *input, float *output, float *temp, const int size, VagueDenoiserContext *s)
285 const int low_size = (size + 1) >> 1;
286 const int high_size = size >> 1;
287 int left_ext = 1, right_ext, i;
290 memcpy(temp + NPAD, input + NPAD, low_size * sizeof(float));
292 right_ext = (size % 2 == 0) ? 2 : 1;
293 symmetric_extension(temp, low_size, left_ext, right_ext);
295 memset(output, 0, (NPAD + NPAD + size) * sizeof(float));
296 findex = (size + 2) >> 1;
298 for (i = 9; i < findex + 11; i++) {
299 const float a = temp[i] * synthesis_low[0];
300 const float b = temp[i] * synthesis_low[1];
301 const float c = temp[i] * synthesis_low[2];
302 const float d = temp[i] * synthesis_low[3];
304 output[2 * i - 13] += a;
305 output[2 * i - 12] += b;
306 output[2 * i - 11] += c;
307 output[2 * i - 10] += d;
308 output[2 * i - 9] += c;
309 output[2 * i - 8] += b;
310 output[2 * i - 7] += a;
313 memcpy(temp + NPAD, input + NPAD + low_size, high_size * sizeof(float));
316 right_ext = (size % 2 == 0) ? 1 : 2;
317 symmetric_extension(temp, high_size, left_ext, right_ext);
319 for (i = 8; i < findex + 11; i++) {
320 const float a = temp[i] * synthesis_high[0];
321 const float b = temp[i] * synthesis_high[1];
322 const float c = temp[i] * synthesis_high[2];
323 const float d = temp[i] * synthesis_high[3];
324 const float e = temp[i] * synthesis_high[4];
326 output[2 * i - 13] += a;
327 output[2 * i - 12] += b;
328 output[2 * i - 11] += c;
329 output[2 * i - 10] += d;
330 output[2 * i - 9] += e;
331 output[2 * i - 8] += d;
332 output[2 * i - 7] += c;
333 output[2 * i - 6] += b;
334 output[2 * i - 5] += a;
338 static void hard_thresholding(float *block, const int width, const int height,
339 const int stride, const float threshold,
342 const float frac = 1.f - percent * 0.01f;
345 for (y = 0; y < height; y++) {
346 for (x = 0; x < width; x++) {
347 if (FFABS(block[x]) <= threshold)
354 static void soft_thresholding(float *block, const int width, const int height, const int stride,
355 const float threshold, const float percent)
357 const float frac = 1.f - percent * 0.01f;
358 const float shift = threshold * 0.01f * percent;
361 for (y = 0; y < height; y++) {
362 for (x = 0; x < width; x++) {
363 const float temp = FFABS(block[x]);
364 if (temp <= threshold)
367 block[x] = (block[x] < 0.f ? -1.f : (block[x] > 0.f ? 1.f : 0.f)) * (temp - shift);
373 static void qian_thresholding(float *block, const int width, const int height,
374 const int stride, const float threshold,
377 const float percent01 = percent * 0.01f;
378 const float tr2 = threshold * threshold * percent01;
379 const float frac = 1.f - percent01;
382 for (y = 0; y < height; y++) {
383 for (x = 0; x < width; x++) {
384 const float temp = FFABS(block[x]);
385 if (temp <= threshold) {
388 const float tp2 = temp * temp;
389 block[x] *= (tp2 - tr2) / tp2;
396 static float bayes_threshold(float *block, const int width, const int height,
397 const int stride, const float threshold)
401 for (int y = 0; y < height; y++) {
402 for (int x = 0; x < width; x++) {
403 mean += block[x] * block[x];
408 mean /= width * height;
410 return threshold * threshold / (FFMAX(sqrtf(mean - threshold), FLT_EPSILON));
413 static void filter(VagueDenoiserContext *s, AVFrame *in, AVFrame *out)
417 for (p = 0; p < s->nb_planes; p++) {
418 const int height = s->planeheight[p];
419 const int width = s->planewidth[p];
420 const uint8_t *srcp8 = in->data[p];
421 const uint16_t *srcp16 = (const uint16_t *)in->data[p];
422 uint8_t *dstp8 = out->data[p];
423 uint16_t *dstp16 = (uint16_t *)out->data[p];
424 float *output = s->block;
425 int h_low_size0 = width;
426 int v_low_size0 = height;
427 int nsteps_transform = s->nsteps;
428 int nsteps_invert = s->nsteps;
429 const float *input = s->block;
431 if (!((1 << p) & s->planes)) {
432 av_image_copy_plane(out->data[p], out->linesize[p], in->data[p], in->linesize[p],
433 s->planewidth[p] * s->bpc, s->planeheight[p]);
438 for (y = 0; y < height; y++) {
439 for (x = 0; x < width; x++)
440 output[x] = srcp8[x];
441 srcp8 += in->linesize[p];
445 for (y = 0; y < height; y++) {
446 for (x = 0; x < width; x++)
447 output[x] = srcp16[x];
448 srcp16 += in->linesize[p] / 2;
453 while (nsteps_transform--) {
454 int low_size = (h_low_size0 + 1) >> 1;
455 float *input = s->block;
456 for (j = 0; j < v_low_size0; j++) {
457 copy(input, s->in + NPAD, h_low_size0);
458 transform_step(s->in, s->out, h_low_size0, low_size, s);
459 copy(s->out + NPAD, input, h_low_size0);
463 low_size = (v_low_size0 + 1) >> 1;
465 for (j = 0; j < h_low_size0; j++) {
466 copyv(input, width, s->in + NPAD, v_low_size0);
467 transform_step(s->in, s->out, v_low_size0, low_size, s);
468 copyh(s->out + NPAD, input, width, v_low_size0);
472 h_low_size0 = (h_low_size0 + 1) >> 1;
473 v_low_size0 = (v_low_size0 + 1) >> 1;
477 s->thresholding(s->block, width, height, width, s->threshold, s->percent);
479 for (int n = 0; n < s->nsteps; n++) {
483 if (n == s->nsteps - 1) {
484 threshold = bayes_threshold(s->block, s->hlowsize[p][n], s->vlowsize[p][n], width, s->threshold);
485 s->thresholding(s->block, s->hlowsize[p][n], s->vlowsize[p][n], width, threshold, s->percent);
487 block = s->block + s->hlowsize[p][n];
488 threshold = bayes_threshold(block, s->hhighsize[p][n], s->vlowsize[p][n], width, s->threshold);
489 s->thresholding(block, s->hhighsize[p][n], s->vlowsize[p][n], width, threshold, s->percent);
490 block = s->block + s->vlowsize[p][n] * width;
491 threshold = bayes_threshold(block, s->hlowsize[p][n], s->vhighsize[p][n], width, s->threshold);
492 s->thresholding(block, s->hlowsize[p][n], s->vhighsize[p][n], width, threshold, s->percent);
493 block = s->block + s->hlowsize[p][n] + s->vlowsize[p][n] * width;
494 threshold = bayes_threshold(block, s->hhighsize[p][n], s->vhighsize[p][n], width, s->threshold);
495 s->thresholding(block, s->hhighsize[p][n], s->vhighsize[p][n], width, threshold, s->percent);
499 while (nsteps_invert--) {
500 const int idx = s->vlowsize[p][nsteps_invert] + s->vhighsize[p][nsteps_invert];
501 const int idx2 = s->hlowsize[p][nsteps_invert] + s->hhighsize[p][nsteps_invert];
502 float * idx3 = s->block;
503 for (i = 0; i < idx2; i++) {
504 copyv(idx3, width, s->in + NPAD, idx);
505 invert_step(s->in, s->out, s->tmp, idx, s);
506 copyh(s->out + NPAD, idx3, width, idx);
511 for (i = 0; i < idx; i++) {
512 copy(idx3, s->in + NPAD, idx2);
513 invert_step(s->in, s->out, s->tmp, idx2, s);
514 copy(s->out + NPAD, idx3, idx2);
520 for (y = 0; y < height; y++) {
521 for (x = 0; x < width; x++)
522 dstp8[x] = av_clip_uint8(input[x] + 0.5f);
524 dstp8 += out->linesize[p];
527 for (y = 0; y < height; y++) {
528 for (x = 0; x < width; x++)
529 dstp16[x] = av_clip(input[x] + 0.5f, 0, s->peak);
531 dstp16 += out->linesize[p] / 2;
537 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
539 AVFilterContext *ctx = inlink->dst;
540 VagueDenoiserContext *s = ctx->priv;
541 AVFilterLink *outlink = ctx->outputs[0];
543 int direct = av_frame_is_writable(in);
548 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
551 return AVERROR(ENOMEM);
554 av_frame_copy_props(out, in);
562 return ff_filter_frame(outlink, out);
565 static av_cold int init(AVFilterContext *ctx)
567 VagueDenoiserContext *s = ctx->priv;
571 s->thresholding = hard_thresholding;
574 s->thresholding = soft_thresholding;
577 s->thresholding = qian_thresholding;
584 static av_cold void uninit(AVFilterContext *ctx)
586 VagueDenoiserContext *s = ctx->priv;
594 static const AVFilterPad vaguedenoiser_inputs[] = {
597 .type = AVMEDIA_TYPE_VIDEO,
598 .config_props = config_input,
599 .filter_frame = filter_frame,
605 static const AVFilterPad vaguedenoiser_outputs[] = {
608 .type = AVMEDIA_TYPE_VIDEO
613 const AVFilter ff_vf_vaguedenoiser = {
614 .name = "vaguedenoiser",
615 .description = NULL_IF_CONFIG_SMALL("Apply a Wavelet based Denoiser."),
616 .priv_size = sizeof(VagueDenoiserContext),
617 .priv_class = &vaguedenoiser_class,
620 .query_formats = query_formats,
621 .inputs = vaguedenoiser_inputs,
622 .outputs = vaguedenoiser_outputs,
623 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,