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 {
61 void (*thresholding)(float *block, const int width, const int height,
62 const int stride, const float threshold,
63 const float percent, const int nsteps);
64 } VagueDenoiserContext;
66 #define OFFSET(x) offsetof(VagueDenoiserContext, x)
67 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_FILTERING_PARAM
68 static const AVOption vaguedenoiser_options[] = {
69 { "threshold", "set filtering strength", OFFSET(threshold), AV_OPT_TYPE_FLOAT, {.dbl=2.}, 0,DBL_MAX, FLAGS },
70 { "method", "set filtering method", OFFSET(method), AV_OPT_TYPE_INT, {.i64=2 }, 0, 2, FLAGS, "method" },
71 { "hard", "hard thresholding", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, FLAGS, "method" },
72 { "soft", "soft thresholding", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, FLAGS, "method" },
73 { "garrote", "garotte thresholding", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, FLAGS, "method" },
74 { "nsteps", "set number of steps", OFFSET(nsteps), AV_OPT_TYPE_INT, {.i64=6 }, 1, 32, FLAGS },
75 { "percent", "set percent of full denoising", OFFSET(percent),AV_OPT_TYPE_FLOAT, {.dbl=85}, 0,100, FLAGS },
76 { "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15 }, 0, 15, FLAGS },
80 AVFILTER_DEFINE_CLASS(vaguedenoiser);
84 static const float analysis_low[9] = {
85 0.037828455506995f, -0.023849465019380f, -0.110624404418423f, 0.377402855612654f,
86 0.852698679009403f, 0.377402855612654f, -0.110624404418423f, -0.023849465019380f, 0.037828455506995f
89 static const float analysis_high[7] = {
90 -0.064538882628938f, 0.040689417609558f, 0.418092273222212f, -0.788485616405664f,
91 0.418092273222212f, 0.040689417609558f, -0.064538882628938f
94 static const float synthesis_low[7] = {
95 -0.064538882628938f, -0.040689417609558f, 0.418092273222212f, 0.788485616405664f,
96 0.418092273222212f, -0.040689417609558f, -0.064538882628938f
99 static const float synthesis_high[9] = {
100 -0.037828455506995f, -0.023849465019380f, 0.110624404418423f, 0.377402855612654f,
101 -0.852698679009403f, 0.377402855612654f, 0.110624404418423f, -0.023849465019380f, -0.037828455506995f
104 static int query_formats(AVFilterContext *ctx)
106 static const enum AVPixelFormat pix_fmts[] = {
107 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10,
108 AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
109 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
110 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
111 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
112 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
113 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
115 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
116 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
117 AV_PIX_FMT_YUV440P10,
118 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
119 AV_PIX_FMT_YUV440P12,
120 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
121 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
122 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
123 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
126 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
128 return AVERROR(ENOMEM);
129 return ff_set_common_formats(ctx, fmts_list);
132 static int config_input(AVFilterLink *inlink)
134 VagueDenoiserContext *s = inlink->dst->priv;
135 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
136 int p, i, nsteps_width, nsteps_height, nsteps_max;
138 s->depth = desc->comp[0].depth;
139 s->bpc = (s->depth + 7) / 8;
140 s->nb_planes = desc->nb_components;
142 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
143 s->planeheight[0] = s->planeheight[3] = inlink->h;
144 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
145 s->planewidth[0] = s->planewidth[3] = inlink->w;
147 s->block = av_malloc_array(inlink->w * inlink->h, sizeof(*s->block));
148 s->in = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->in));
149 s->out = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->out));
150 s->tmp = av_malloc_array(32 + FFMAX(inlink->w, inlink->h), sizeof(*s->tmp));
152 if (!s->block || !s->in || !s->out || !s->tmp)
153 return AVERROR(ENOMEM);
155 s->threshold *= 1 << (s->depth - 8);
156 s->peak = (1 << s->depth) - 1;
158 nsteps_width = ((s->planes & 2 || s->planes & 4) && s->nb_planes > 1) ? s->planewidth[1] : s->planewidth[0];
159 nsteps_height = ((s->planes & 2 || s->planes & 4) && s->nb_planes > 1) ? s->planeheight[1] : s->planeheight[0];
161 for (nsteps_max = 1; nsteps_max < 15; nsteps_max++) {
162 if (pow(2, nsteps_max) >= nsteps_width || pow(2, nsteps_max) >= nsteps_height)
166 s->nsteps = FFMIN(s->nsteps, nsteps_max - 2);
168 for (p = 0; p < 4; p++) {
169 s->hlowsize[p][0] = (s->planewidth[p] + 1) >> 1;
170 s->hhighsize[p][0] = s->planewidth[p] >> 1;
171 s->vlowsize[p][0] = (s->planeheight[p] + 1) >> 1;
172 s->vhighsize[p][0] = s->planeheight[p] >> 1;
174 for (i = 1; i < s->nsteps; i++) {
175 s->hlowsize[p][i] = (s->hlowsize[p][i - 1] + 1) >> 1;
176 s->hhighsize[p][i] = s->hlowsize[p][i - 1] >> 1;
177 s->vlowsize[p][i] = (s->vlowsize[p][i - 1] + 1) >> 1;
178 s->vhighsize[p][i] = s->vlowsize[p][i - 1] >> 1;
185 static inline void copy(const float *p1, float *p2, const int length)
187 memcpy(p2, p1, length * sizeof(float));
190 static inline void copyv(const float *p1, const int stride1, float *p2, const int length)
194 for (i = 0; i < length; i++) {
200 static inline void copyh(const float *p1, float *p2, const int stride2, const int length)
204 for (i = 0; i < length; i++) {
210 // Do symmetric extension of data using prescribed symmetries
211 // Original values are in output[npad] through output[npad+size-1]
212 // 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)
213 // extension at left bdry is ... 3 2 1 0 | 0 1 2 3 ...
214 // same for right boundary
215 // if right_ext=1 then ... 3 2 1 0 | 1 2 3
216 static void symmetric_extension(float *output, const int size, const int left_ext, const int right_ext)
219 int last = NPAD - 1 + size;
220 const int originalLast = last;
224 output[--first] = output[NPAD];
226 output[++last] = output[originalLast];
230 for (i = 0; i < nextend; i++)
231 output[--first] = output[NPAD + 1 + i];
233 idx = NPAD + NPAD - 1 + size;
236 nextend = idx - last;
237 for (i = 0; i < nextend; i++)
238 output[++last] = output[originalLast - 1 - i];
241 static void transform_step(float *input, float *output, const int size, const int low_size, VagueDenoiserContext *s)
245 symmetric_extension(input, size, 1, 1);
247 for (i = NPAD; i < NPAD + low_size; i++) {
248 const float a = input[2 * i - 14] * analysis_low[0];
249 const float b = input[2 * i - 13] * analysis_low[1];
250 const float c = input[2 * i - 12] * analysis_low[2];
251 const float d = input[2 * i - 11] * analysis_low[3];
252 const float e = input[2 * i - 10] * analysis_low[4];
253 const float f = input[2 * i - 9] * analysis_low[3];
254 const float g = input[2 * i - 8] * analysis_low[2];
255 const float h = input[2 * i - 7] * analysis_low[1];
256 const float k = input[2 * i - 6] * analysis_low[0];
258 output[i] = a + b + c + d + e + f + g + h + k;
261 for (i = NPAD; i < NPAD + low_size; i++) {
262 const float a = input[2 * i - 12] * analysis_high[0];
263 const float b = input[2 * i - 11] * analysis_high[1];
264 const float c = input[2 * i - 10] * analysis_high[2];
265 const float d = input[2 * i - 9] * analysis_high[3];
266 const float e = input[2 * i - 8] * analysis_high[2];
267 const float f = input[2 * i - 7] * analysis_high[1];
268 const float g = input[2 * i - 6] * analysis_high[0];
270 output[i + low_size] = a + b + c + d + e + f + g;
274 static void invert_step(const float *input, float *output, float *temp, const int size, VagueDenoiserContext *s)
276 const int low_size = (size + 1) >> 1;
277 const int high_size = size >> 1;
278 int left_ext = 1, right_ext, i;
281 memcpy(temp + NPAD, input + NPAD, low_size * sizeof(float));
283 right_ext = (size % 2 == 0) ? 2 : 1;
284 symmetric_extension(temp, low_size, left_ext, right_ext);
286 memset(output, 0, (NPAD + NPAD + size) * sizeof(float));
287 findex = (size + 2) >> 1;
289 for (i = 9; i < findex + 11; i++) {
290 const float a = temp[i] * synthesis_low[0];
291 const float b = temp[i] * synthesis_low[1];
292 const float c = temp[i] * synthesis_low[2];
293 const float d = temp[i] * synthesis_low[3];
295 output[2 * i - 13] += a;
296 output[2 * i - 12] += b;
297 output[2 * i - 11] += c;
298 output[2 * i - 10] += d;
299 output[2 * i - 9] += c;
300 output[2 * i - 8] += b;
301 output[2 * i - 7] += a;
304 memcpy(temp + NPAD, input + NPAD + low_size, high_size * sizeof(float));
307 right_ext = (size % 2 == 0) ? 1 : 2;
308 symmetric_extension(temp, high_size, left_ext, right_ext);
310 for (i = 8; i < findex + 11; i++) {
311 const float a = temp[i] * synthesis_high[0];
312 const float b = temp[i] * synthesis_high[1];
313 const float c = temp[i] * synthesis_high[2];
314 const float d = temp[i] * synthesis_high[3];
315 const float e = temp[i] * synthesis_high[4];
317 output[2 * i - 13] += a;
318 output[2 * i - 12] += b;
319 output[2 * i - 11] += c;
320 output[2 * i - 10] += d;
321 output[2 * i - 9] += e;
322 output[2 * i - 8] += d;
323 output[2 * i - 7] += c;
324 output[2 * i - 6] += b;
325 output[2 * i - 5] += a;
329 static void hard_thresholding(float *block, const int width, const int height,
330 const int stride, const float threshold,
331 const float percent, const int unused)
333 const float frac = 1.f - percent * 0.01f;
336 for (y = 0; y < height; y++) {
337 for (x = 0; x < width; x++) {
338 if (FFABS(block[x]) <= threshold)
345 static void soft_thresholding(float *block, const int width, const int height, const int stride,
346 const float threshold, const float percent, const int nsteps)
348 const float frac = 1.f - percent * 0.01f;
349 const float shift = threshold * 0.01f * percent;
354 for (l = 0; l < nsteps; l++) {
359 for (y = 0; y < height; y++) {
360 const int x0 = (y < h) ? w : 0;
361 for (x = x0; x < width; x++) {
362 const float temp = FFABS(block[x]);
363 if (temp <= threshold)
366 block[x] = (block[x] < 0.f ? -1.f : (block[x] > 0.f ? 1.f : 0.f)) * (temp - shift);
372 static void qian_thresholding(float *block, const int width, const int height,
373 const int stride, const float threshold,
374 const float percent, const int unused)
376 const float percent01 = percent * 0.01f;
377 const float tr2 = threshold * threshold * percent01;
378 const float frac = 1.f - percent01;
381 for (y = 0; y < height; y++) {
382 for (x = 0; x < width; x++) {
383 const float temp = FFABS(block[x]);
384 if (temp <= threshold) {
387 const float tp2 = temp * temp;
388 block[x] *= (tp2 - tr2) / tp2;
395 static void filter(VagueDenoiserContext *s, AVFrame *in, AVFrame *out)
399 for (p = 0; p < s->nb_planes; p++) {
400 const int height = s->planeheight[p];
401 const int width = s->planewidth[p];
402 const uint8_t *srcp8 = in->data[p];
403 const uint16_t *srcp16 = (const uint16_t *)in->data[p];
404 uint8_t *dstp8 = out->data[p];
405 uint16_t *dstp16 = (uint16_t *)out->data[p];
406 float *output = s->block;
407 int h_low_size0 = width;
408 int v_low_size0 = height;
409 int nsteps_transform = s->nsteps;
410 int nsteps_invert = s->nsteps;
411 const float *input = s->block;
413 if (!((1 << p) & s->planes)) {
414 av_image_copy_plane(out->data[p], out->linesize[p], in->data[p], in->linesize[p],
415 s->planewidth[p] * s->bpc, s->planeheight[p]);
420 for (y = 0; y < height; y++) {
421 for (x = 0; x < width; x++)
422 output[x] = srcp8[x];
423 srcp8 += in->linesize[p];
427 for (y = 0; y < height; y++) {
428 for (x = 0; x < width; x++)
429 output[x] = srcp16[x];
430 srcp16 += in->linesize[p] / 2;
435 while (nsteps_transform--) {
436 int low_size = (h_low_size0 + 1) >> 1;
437 float *input = s->block;
438 for (j = 0; j < v_low_size0; j++) {
439 copy(input, s->in + NPAD, h_low_size0);
440 transform_step(s->in, s->out, h_low_size0, low_size, s);
441 copy(s->out + NPAD, input, h_low_size0);
445 low_size = (v_low_size0 + 1) >> 1;
447 for (j = 0; j < h_low_size0; j++) {
448 copyv(input, width, s->in + NPAD, v_low_size0);
449 transform_step(s->in, s->out, v_low_size0, low_size, s);
450 copyh(s->out + NPAD, input, width, v_low_size0);
454 h_low_size0 = (h_low_size0 + 1) >> 1;
455 v_low_size0 = (v_low_size0 + 1) >> 1;
458 s->thresholding(s->block, width, height, width, s->threshold, s->percent, s->nsteps);
460 while (nsteps_invert--) {
461 const int idx = s->vlowsize[p][nsteps_invert] + s->vhighsize[p][nsteps_invert];
462 const int idx2 = s->hlowsize[p][nsteps_invert] + s->hhighsize[p][nsteps_invert];
463 float * idx3 = s->block;
464 for (i = 0; i < idx2; i++) {
465 copyv(idx3, width, s->in + NPAD, idx);
466 invert_step(s->in, s->out, s->tmp, idx, s);
467 copyh(s->out + NPAD, idx3, width, idx);
472 for (i = 0; i < idx; i++) {
473 copy(idx3, s->in + NPAD, idx2);
474 invert_step(s->in, s->out, s->tmp, idx2, s);
475 copy(s->out + NPAD, idx3, idx2);
481 for (y = 0; y < height; y++) {
482 for (x = 0; x < width; x++)
483 dstp8[x] = av_clip_uint8(input[x] + 0.5f);
485 dstp8 += out->linesize[p];
488 for (y = 0; y < height; y++) {
489 for (x = 0; x < width; x++)
490 dstp16[x] = av_clip(input[x] + 0.5f, 0, s->peak);
492 dstp16 += out->linesize[p] / 2;
498 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
500 AVFilterContext *ctx = inlink->dst;
501 VagueDenoiserContext *s = ctx->priv;
502 AVFilterLink *outlink = ctx->outputs[0];
504 int direct = av_frame_is_writable(in);
509 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
512 return AVERROR(ENOMEM);
515 av_frame_copy_props(out, in);
523 return ff_filter_frame(outlink, out);
526 static av_cold int init(AVFilterContext *ctx)
528 VagueDenoiserContext *s = ctx->priv;
532 s->thresholding = hard_thresholding;
535 s->thresholding = soft_thresholding;
538 s->thresholding = qian_thresholding;
545 static av_cold void uninit(AVFilterContext *ctx)
547 VagueDenoiserContext *s = ctx->priv;
555 static const AVFilterPad vaguedenoiser_inputs[] = {
558 .type = AVMEDIA_TYPE_VIDEO,
559 .config_props = config_input,
560 .filter_frame = filter_frame,
566 static const AVFilterPad vaguedenoiser_outputs[] = {
569 .type = AVMEDIA_TYPE_VIDEO
574 AVFilter ff_vf_vaguedenoiser = {
575 .name = "vaguedenoiser",
576 .description = NULL_IF_CONFIG_SMALL("Apply a Wavelet based Denoiser."),
577 .priv_size = sizeof(VagueDenoiserContext),
578 .priv_class = &vaguedenoiser_class,
581 .query_formats = query_formats,
582 .inputs = vaguedenoiser_inputs,
583 .outputs = vaguedenoiser_outputs,
584 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,