FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavfilter/vf_nlmeans.c
Date: 2026-09-25 17:25:24
Exec Total Coverage
Lines: 0 170 0.0%
Functions: 0 9 0.0%
Branches: 0 66 0.0%
Line Branch Exec Source
1 /*
2 * Copyright (c) 2016 Clément Bœsch <u pkh me>
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
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 GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 /**
22 * @todo
23 * - better automatic defaults? see "Parameters" @ http://www.ipol.im/pub/art/2011/bcm_nlm/
24 * - temporal support (probably doesn't need any displacement according to
25 * "Denoising image sequences does not require motion estimation")
26 * - Bayer pixel format support for at least raw photos? (DNG support would be
27 * handy here)
28 * - FATE test (probably needs visual threshold test mechanism due to the use
29 * of floats)
30 */
31
32 #include "libavutil/avassert.h"
33 #include "libavutil/mem.h"
34 #include "libavutil/opt.h"
35 #include "libavutil/pixdesc.h"
36 #include "avfilter.h"
37 #include "filters.h"
38 #include "vf_nlmeans.h"
39 #include "vf_nlmeans_init.h"
40 #include "video.h"
41
42 typedef struct NLMeansContext {
43 const AVClass *class;
44 int nb_planes;
45 int chroma_w, chroma_h;
46 double pdiff_scale; // invert of the filtering parameter (sigma*10) squared
47 double sigma; // denoising strength
48 int patch_size, patch_hsize; // patch size and half size
49 int patch_size_uv, patch_hsize_uv; // patch size and half size for chroma planes
50 int research_size, research_hsize; // research size and half size
51 int research_size_uv, research_hsize_uv; // research size and half size for chroma planes
52 uint32_t *ii_orig; // integral image
53 uint32_t *ii; // integral image starting after the 0-line and 0-column
54 int ii_w, ii_h; // width and height of the integral image
55 ptrdiff_t ii_lz_32; // linesize in 32-bit units of the integral image
56 float *total_weight; // total weight for every pixel
57 float *sum; // weighted sum for every pixel
58 int linesize; // sum and total_weight linesize
59 float *weight_lut; // lookup table mapping (scaled) patch differences to their associated weights
60 uint32_t max_meaningful_diff; // maximum difference considered (if the patch difference is too high we ignore the pixel)
61 NLMeansDSPContext dsp;
62 } NLMeansContext;
63
64 #define OFFSET(x) offsetof(NLMeansContext, x)
65 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
66 static const AVOption nlmeans_options[] = {
67 { "s", "denoising strength", OFFSET(sigma), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 }, 1.0, 30.0, FLAGS },
68 { "p", "patch size", OFFSET(patch_size), AV_OPT_TYPE_INT, { .i64 = 3*2+1 }, 0, 99, FLAGS },
69 { "pc", "patch size for chroma planes", OFFSET(patch_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS },
70 { "r", "research window", OFFSET(research_size), AV_OPT_TYPE_INT, { .i64 = 7*2+1 }, 0, 99, FLAGS },
71 { "rc", "research window for chroma planes", OFFSET(research_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS },
72 { NULL }
73 };
74
75 AVFILTER_DEFINE_CLASS(nlmeans);
76
77 static const enum AVPixelFormat pix_fmts[] = {
78 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
79 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
80 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
81 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
82 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
83 AV_PIX_FMT_YUVJ411P,
84 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
85 AV_PIX_FMT_NONE
86 };
87
88 /**
89 * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
90 * be readable).
91 *
92 * On the other hand, the line above dst and the column to its left are always
93 * readable.
94 *
95 * There is little point in having this function SIMDified as it is likely too
96 * complex and only handle small portions of the image.
97 *
98 * @param dst integral image
99 * @param dst_linesize_32 integral image linesize (in 32-bit integers unit)
100 * @param startx integral starting x position
101 * @param starty integral starting y position
102 * @param src source plane buffer
103 * @param linesize source plane linesize
104 * @param offx source offsetting in x
105 * @param offy source offsetting in y
106 * @paran r absolute maximum source offsetting
107 * @param sw source width
108 * @param sh source height
109 * @param w width to compute
110 * @param h height to compute
111 */
112 ✗ static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t dst_linesize_32,
113 int startx, int starty,
114 const uint8_t *src, ptrdiff_t linesize,
115 int offx, int offy, int r, int sw, int sh,
116 int w, int h)
117 {
118 ✗ for (int y = starty; y < starty + h; y++) {
119 ✗ uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
120 ✗ const int s1y = av_clip(y - r, 0, sh - 1);
121 ✗ const int s2y = av_clip(y - (r + offy), 0, sh - 1);
122
123 ✗ for (int x = startx; x < startx + w; x++) {
124 ✗ const int s1x = av_clip(x - r, 0, sw - 1);
125 ✗ const int s2x = av_clip(x - (r + offx), 0, sw - 1);
126 ✗ const uint8_t v1 = src[s1y*linesize + s1x];
127 ✗ const uint8_t v2 = src[s2y*linesize + s2x];
128 ✗ const int d = v1 - v2;
129 ✗ acc += d * d;
130 ✗ dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
131 }
132 }
133 ✗ }
134
135 /*
136 * Compute the sum of squared difference integral image
137 * http://www.ipol.im/pub/art/2014/57/
138 * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
139 *
140 * @param ii integral image of dimension (w+e*2) x (h+e*2) with
141 * an additional zeroed top line and column already
142 * "applied" to the pointer value
143 * @param ii_linesize_32 integral image linesize (in 32-bit integers unit)
144 * @param src source plane buffer
145 * @param linesize source plane linesize
146 * @param offx x-offsetting ranging in [-e;e]
147 * @param offy y-offsetting ranging in [-e;e]
148 * @param w source width
149 * @param h source height
150 * @param e research padding edge
151 */
152 ✗ static void compute_ssd_integral_image(const NLMeansDSPContext *dsp,
153 uint32_t *ii, ptrdiff_t ii_linesize_32,
154 const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
155 int e, int w, int h)
156 {
157 // ii has a surrounding padding of thickness "e"
158 ✗ const int ii_w = w + e*2;
159 ✗ const int ii_h = h + e*2;
160
161 // we center the first source
162 ✗ const int s1x = e;
163 ✗ const int s1y = e;
164
165 // 2nd source is the frame with offsetting
166 ✗ const int s2x = e + offx;
167 ✗ const int s2y = e + offy;
168
169 // get the dimension of the overlapping rectangle where it is always safe
170 // to compare the 2 sources pixels
171 ✗ const int startx_safe = FFMAX(s1x, s2x);
172 ✗ const int starty_safe = FFMAX(s1y, s2y);
173 ✗ const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
174 ✗ const int endy_safe = FFMIN(s1y + h, s2y + h);
175
176 // deduce the safe area width and height
177 ✗ const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
178 ✗ const int safe_ph = endy_safe - starty_safe;
179
180 // adjusted end x position of the safe area after width of the safe area gets aligned
181 ✗ const int endx_safe = startx_safe + safe_pw;
182
183 // top part where only one of s1 and s2 is still readable, or none at all
184 ✗ compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
185 0, 0,
186 src, linesize,
187 offx, offy, e, w, h,
188 ii_w, starty_safe);
189
190 // fill the left column integral required to compute the central
191 // overlapping one
192 ✗ compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
193 0, starty_safe,
194 src, linesize,
195 offx, offy, e, w, h,
196 startx_safe, safe_ph);
197
198 // main and safe part of the integral
199 av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
200 av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
201 av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
202 av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
203 ✗ if (safe_pw && safe_ph)
204 ✗ dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
205 ✗ src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
206 ✗ src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
207 safe_pw, safe_ph);
208
209 // right part of the integral
210 ✗ compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
211 endx_safe, starty_safe,
212 src, linesize,
213 offx, offy, e, w, h,
214 ii_w - endx_safe, safe_ph);
215
216 // bottom part where only one of s1 and s2 is still readable, or none at all
217 ✗ compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
218 0, endy_safe,
219 src, linesize,
220 offx, offy, e, w, h,
221 ii_w, ii_h - endy_safe);
222 ✗ }
223
224 ✗ static int config_input(AVFilterLink *inlink)
225 {
226 ✗ AVFilterContext *ctx = inlink->dst;
227 ✗ NLMeansContext *s = ctx->priv;
228 ✗ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
229 ✗ const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
230 ✗ + FFMAX(s->patch_hsize, s->patch_hsize_uv);
231
232 ✗ s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
233 ✗ s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
234 ✗ s->nb_planes = av_pix_fmt_count_planes(inlink->format);
235
236 /* Allocate the integral image with extra edges of thickness "e"
237 *
238 * +_+-------------------------------+
239 * |0|0000000000000000000000000000000|
240 * +-x-------------------------------+
241 * |0|\ ^ |
242 * |0| ii | e |
243 * |0| v |
244 * |0| +-----------------------+ |
245 * |0| | | |
246 * |0|<->| | |
247 * |0| e | | |
248 * |0| | | |
249 * |0| +-----------------------+ |
250 * |0| |
251 * |0| |
252 * |0| |
253 * +-+-------------------------------+
254 */
255 ✗ s->ii_w = inlink->w + e*2;
256 ✗ s->ii_h = inlink->h + e*2;
257
258 // align to 4 the linesize, "+1" is for the space of the left 0-column
259 ✗ s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
260
261 // "+1" is for the space of the top 0-line
262 ✗ s->ii_orig = av_calloc(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
263 ✗ if (!s->ii_orig)
264 ✗ return AVERROR(ENOMEM);
265
266 // skip top 0-line and left 0-column
267 ✗ s->ii = s->ii_orig + s->ii_lz_32 + 1;
268
269 // allocate weighted average for every pixel
270 ✗ s->linesize = inlink->w + 100;
271 ✗ s->total_weight = av_malloc_array(s->linesize, inlink->h * sizeof(*s->total_weight));
272 ✗ s->sum = av_malloc_array(s->linesize, inlink->h * sizeof(*s->sum));
273 ✗ if (!s->total_weight || !s->sum)
274 ✗ return AVERROR(ENOMEM);
275
276 ✗ return 0;
277 }
278
279 struct thread_data {
280 const uint8_t *src;
281 ptrdiff_t src_linesize;
282 int startx, starty;
283 int endx, endy;
284 const uint32_t *ii_start;
285 int p;
286 };
287
288 ✗ static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
289 {
290 ✗ NLMeansContext *s = ctx->priv;
291 ✗ const uint32_t max_meaningful_diff = s->max_meaningful_diff;
292 ✗ const struct thread_data *td = arg;
293 ✗ const ptrdiff_t src_linesize = td->src_linesize;
294 ✗ const int process_h = td->endy - td->starty;
295 ✗ const int slice_start = ff_slice_pos(process_h, jobnr, nb_jobs);
296 ✗ const int slice_end = ff_slice_pos(process_h, jobnr + 1, nb_jobs);
297 ✗ const int starty = td->starty + slice_start;
298 ✗ const int endy = td->starty + slice_end;
299 ✗ const int p = td->p;
300 ✗ const uint32_t *ii = td->ii_start + (starty - p - 1) * s->ii_lz_32 - p - 1;
301 ✗ const int dist_b = 2*p + 1;
302 ✗ const int dist_d = dist_b * s->ii_lz_32;
303 ✗ const int dist_e = dist_d + dist_b;
304 ✗ const float *const weight_lut = s->weight_lut;
305 ✗ NLMeansDSPContext *dsp = &s->dsp;
306
307 ✗ for (int y = starty; y < endy; y++) {
308 ✗ const uint8_t *const src = td->src + y*src_linesize;
309 ✗ float *total_weight = s->total_weight + y*s->linesize;
310 ✗ float *sum = s->sum + y*s->linesize;
311 ✗ const uint32_t *const iia = ii;
312 ✗ const uint32_t *const iib = ii + dist_b;
313 ✗ const uint32_t *const iid = ii + dist_d;
314 ✗ const uint32_t *const iie = ii + dist_e;
315
316 ✗ dsp->compute_weights_line(iia, iib, iid, iie, src, total_weight, sum,
317 weight_lut, max_meaningful_diff,
318 ✗ td->startx, td->endx);
319 ✗ ii += s->ii_lz_32;
320 }
321 ✗ return 0;
322 }
323
324 ✗ static void weight_averages(uint8_t *dst, ptrdiff_t dst_linesize,
325 const uint8_t *src, ptrdiff_t src_linesize,
326 float *total_weight, float *sum, ptrdiff_t linesize,
327 int w, int h)
328 {
329 ✗ for (int y = 0; y < h; y++) {
330 ✗ for (int x = 0; x < w; x++) {
331 // Also weight the centered pixel
332 ✗ total_weight[x] += 1.f;
333 ✗ sum[x] += 1.f * src[x];
334 ✗ dst[x] = av_clip_uint8(sum[x] / total_weight[x] + 0.5f);
335 }
336 ✗ dst += dst_linesize;
337 ✗ src += src_linesize;
338 ✗ total_weight += linesize;
339 ✗ sum += linesize;
340 }
341 ✗ }
342
343 ✗ static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
344 uint8_t *dst, ptrdiff_t dst_linesize,
345 const uint8_t *src, ptrdiff_t src_linesize)
346 {
347 ✗ NLMeansContext *s = ctx->priv;
348 /* patches center points cover the whole research window so the patches
349 * themselves overflow the research window */
350 ✗ const int e = r + p;
351 /* focus an integral pointer on the centered image (s1) */
352 ✗ const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;
353
354 ✗ memset(s->total_weight, 0, s->linesize * h * sizeof(*s->total_weight));
355 ✗ memset(s->sum, 0, s->linesize * h * sizeof(*s->sum));
356
357 ✗ for (int offy = -r; offy <= r; offy++) {
358 ✗ for (int offx = -r; offx <= r; offx++) {
359 ✗ if (offx || offy) {
360 ✗ struct thread_data td = {
361 ✗ .src = src + offy*src_linesize + offx,
362 .src_linesize = src_linesize,
363 ✗ .startx = FFMAX(0, -offx),
364 ✗ .starty = FFMAX(0, -offy),
365 ✗ .endx = FFMIN(w, w - offx),
366 ✗ .endy = FFMIN(h, h - offy),
367 ✗ .ii_start = centered_ii + offy*s->ii_lz_32 + offx,
368 .p = p,
369 };
370
371 ✗ compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32,
372 src, src_linesize,
373 offx, offy, e, w, h);
374 ✗ ff_filter_execute(ctx, nlmeans_slice, &td, NULL,
375 ✗ FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
376 }
377 }
378 }
379
380 ✗ weight_averages(dst, dst_linesize, src, src_linesize,
381 ✗ s->total_weight, s->sum, s->linesize, w, h);
382
383 ✗ return 0;
384 }
385
386 ✗ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
387 {
388 ✗ AVFilterContext *ctx = inlink->dst;
389 ✗ NLMeansContext *s = ctx->priv;
390 ✗ AVFilterLink *outlink = ctx->outputs[0];
391
392 ✗ AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
393 ✗ if (!out) {
394 ✗ av_frame_free(&in);
395 ✗ return AVERROR(ENOMEM);
396 }
397 ✗ av_frame_copy_props(out, in);
398
399 ✗ for (int i = 0; i < s->nb_planes; i++) {
400 ✗ const int w = i ? s->chroma_w : inlink->w;
401 ✗ const int h = i ? s->chroma_h : inlink->h;
402 ✗ const int p = i ? s->patch_hsize_uv : s->patch_hsize;
403 ✗ const int r = i ? s->research_hsize_uv : s->research_hsize;
404 ✗ nlmeans_plane(ctx, w, h, p, r,
405 ✗ out->data[i], out->linesize[i],
406 ✗ in->data[i], in->linesize[i]);
407 }
408
409 ✗ av_frame_free(&in);
410 ✗ return ff_filter_frame(outlink, out);
411 }
412
413 #define CHECK_ODD_FIELD(field, name) do { \
414 if (!(s->field & 1)) { \
415 s->field |= 1; \
416 av_log(ctx, AV_LOG_WARNING, name " size must be odd, " \
417 "setting it to %d\n", s->field); \
418 } \
419 } while (0)
420
421 ✗ static av_cold int init(AVFilterContext *ctx)
422 {
423 ✗ NLMeansContext *s = ctx->priv;
424 ✗ const double h = s->sigma * 10.;
425
426 ✗ s->pdiff_scale = 1. / (h * h);
427 ✗ s->max_meaningful_diff = log(255.) / s->pdiff_scale;
428 ✗ s->weight_lut = av_calloc(s->max_meaningful_diff + 1, sizeof(*s->weight_lut));
429 ✗ if (!s->weight_lut)
430 ✗ return AVERROR(ENOMEM);
431 ✗ for (int i = 0; i < s->max_meaningful_diff; i++)
432 ✗ s->weight_lut[i] = exp(-i * s->pdiff_scale);
433
434 ✗ CHECK_ODD_FIELD(research_size, "Luma research window");
435 ✗ CHECK_ODD_FIELD(patch_size, "Luma patch");
436
437 ✗ if (!s->research_size_uv) s->research_size_uv = s->research_size;
438 ✗ if (!s->patch_size_uv) s->patch_size_uv = s->patch_size;
439
440 ✗ CHECK_ODD_FIELD(research_size_uv, "Chroma research window");
441 ✗ CHECK_ODD_FIELD(patch_size_uv, "Chroma patch");
442
443 ✗ s->research_hsize = s->research_size / 2;
444 ✗ s->research_hsize_uv = s->research_size_uv / 2;
445 ✗ s->patch_hsize = s->patch_size / 2;
446 ✗ s->patch_hsize_uv = s->patch_size_uv / 2;
447
448 ✗ av_log(ctx, AV_LOG_DEBUG, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
449 s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
450 s->patch_size, s->patch_size, s->patch_size_uv, s->patch_size_uv);
451
452 ✗ ff_nlmeans_init(&s->dsp);
453
454 ✗ return 0;
455 }
456
457 ✗ static av_cold void uninit(AVFilterContext *ctx)
458 {
459 ✗ NLMeansContext *s = ctx->priv;
460 ✗ av_freep(&s->weight_lut);
461 ✗ av_freep(&s->ii_orig);
462 ✗ av_freep(&s->total_weight);
463 ✗ av_freep(&s->sum);
464 ✗ }
465
466 static const AVFilterPad nlmeans_inputs[] = {
467 {
468 .name = "default",
469 .type = AVMEDIA_TYPE_VIDEO,
470 .config_props = config_input,
471 .filter_frame = filter_frame,
472 },
473 };
474
475 const FFFilter ff_vf_nlmeans = {
476 .p.name = "nlmeans",
477 .p.description = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
478 .p.priv_class = &nlmeans_class,
479 .p.flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,
480 .priv_size = sizeof(NLMeansContext),
481 .init = init,
482 .uninit = uninit,
483 FILTER_INPUTS(nlmeans_inputs),
484 FILTER_OUTPUTS(ff_video_default_filterpad),
485 FILTER_PIXFMTS_ARRAY(pix_fmts),
486 };
487