2 * Copyright (c) 2016 Clément Bœsch <u pkh me>
4 * This file is part of FFmpeg.
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.
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.
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
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
28 * - FATE test (probably needs visual threshold test mechanism due to the use
32 #include "libavutil/avassert.h"
33 #include "libavutil/opt.h"
34 #include "libavutil/pixdesc.h"
38 #include "vf_nlmeans.h"
47 * Note: WEIGHT_LUT_SIZE must be larger than max_meaningful_diff
48 * (log(255)*max(h)^2, which is approximately 500000 with the current
49 * maximum sigma of 30).
51 #define WEIGHT_LUT_SIZE 500000
53 typedef struct NLMeansContext {
56 int chroma_w, chroma_h;
57 double pdiff_scale; // invert of the filtering parameter (sigma*10) squared
58 double sigma; // denoising strength
59 int patch_size, patch_hsize; // patch size and half size
60 int patch_size_uv, patch_hsize_uv; // patch size and half size for chroma planes
61 int research_size, research_hsize; // research size and half size
62 int research_size_uv, research_hsize_uv; // research size and half size for chroma planes
63 uint32_t *ii_orig; // integral image
64 uint32_t *ii; // integral image starting after the 0-line and 0-column
65 int ii_w, ii_h; // width and height of the integral image
66 ptrdiff_t ii_lz_32; // linesize in 32-bit units of the integral image
67 struct weighted_avg *wa; // weighted average of every pixel
68 ptrdiff_t wa_linesize; // linesize for wa in struct size unit
69 float weight_lut[WEIGHT_LUT_SIZE]; // lookup table mapping (scaled) patch differences to their associated weights
70 uint32_t max_meaningful_diff; // maximum difference considered (if the patch difference is too high we ignore the pixel)
71 NLMeansDSPContext dsp;
74 #define OFFSET(x) offsetof(NLMeansContext, x)
75 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
76 static const AVOption nlmeans_options[] = {
77 { "s", "denoising strength", OFFSET(sigma), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 }, 1.0, 30.0, FLAGS },
78 { "p", "patch size", OFFSET(patch_size), AV_OPT_TYPE_INT, { .i64 = 3*2+1 }, 0, 99, FLAGS },
79 { "pc", "patch size for chroma planes", OFFSET(patch_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS },
80 { "r", "research window", OFFSET(research_size), AV_OPT_TYPE_INT, { .i64 = 7*2+1 }, 0, 99, FLAGS },
81 { "rc", "research window for chroma planes", OFFSET(research_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS },
85 AVFILTER_DEFINE_CLASS(nlmeans);
87 static int query_formats(AVFilterContext *ctx)
89 static const enum AVPixelFormat pix_fmts[] = {
90 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
91 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
92 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
93 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
94 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
96 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
100 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
102 return AVERROR(ENOMEM);
103 return ff_set_common_formats(ctx, fmts_list);
107 * Compute squared difference of the safe area (the zone where s1 and s2
108 * overlap). It is likely the largest integral zone, so it is interesting to do
109 * as little checks as possible; contrary to the unsafe version of this
110 * function, we do not need any clipping here.
112 * The line above dst and the column to its left are always readable.
114 static void compute_safe_ssd_integral_image_c(uint32_t *dst, ptrdiff_t dst_linesize_32,
115 const uint8_t *s1, ptrdiff_t linesize1,
116 const uint8_t *s2, ptrdiff_t linesize2,
120 const uint32_t *dst_top = dst - dst_linesize_32;
122 /* SIMD-friendly assumptions allowed here */
123 av_assert2(!(w & 0xf) && w >= 16 && h >= 1);
125 for (y = 0; y < h; y++) {
126 for (x = 0; x < w; x += 4) {
127 const int d0 = s1[x ] - s2[x ];
128 const int d1 = s1[x + 1] - s2[x + 1];
129 const int d2 = s1[x + 2] - s2[x + 2];
130 const int d3 = s1[x + 3] - s2[x + 3];
132 dst[x ] = dst_top[x ] - dst_top[x - 1] + d0*d0;
133 dst[x + 1] = dst_top[x + 1] - dst_top[x ] + d1*d1;
134 dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
135 dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;
137 dst[x ] += dst[x - 1];
138 dst[x + 1] += dst[x ];
139 dst[x + 2] += dst[x + 1];
140 dst[x + 3] += dst[x + 2];
144 dst += dst_linesize_32;
145 dst_top += dst_linesize_32;
150 * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
153 * On the other hand, the line above dst and the column to its left are always
156 * There is little point in having this function SIMDified as it is likely too
157 * complex and only handle small portions of the image.
159 * @param dst integral image
160 * @param dst_linesize_32 integral image linesize (in 32-bit integers unit)
161 * @param startx integral starting x position
162 * @param starty integral starting y position
163 * @param src source plane buffer
164 * @param linesize source plane linesize
165 * @param offx source offsetting in x
166 * @param offy source offsetting in y
167 * @paran r absolute maximum source offsetting
168 * @param sw source width
169 * @param sh source height
170 * @param w width to compute
171 * @param h height to compute
173 static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t dst_linesize_32,
174 int startx, int starty,
175 const uint8_t *src, ptrdiff_t linesize,
176 int offx, int offy, int r, int sw, int sh,
181 for (y = starty; y < starty + h; y++) {
182 uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
183 const int s1y = av_clip(y - r, 0, sh - 1);
184 const int s2y = av_clip(y - (r + offy), 0, sh - 1);
186 for (x = startx; x < startx + w; x++) {
187 const int s1x = av_clip(x - r, 0, sw - 1);
188 const int s2x = av_clip(x - (r + offx), 0, sw - 1);
189 const uint8_t v1 = src[s1y*linesize + s1x];
190 const uint8_t v2 = src[s2y*linesize + s2x];
191 const int d = v1 - v2;
193 dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
199 * Compute the sum of squared difference integral image
200 * http://www.ipol.im/pub/art/2014/57/
201 * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
203 * @param ii integral image of dimension (w+e*2) x (h+e*2) with
204 * an additional zeroed top line and column already
205 * "applied" to the pointer value
206 * @param ii_linesize_32 integral image linesize (in 32-bit integers unit)
207 * @param src source plane buffer
208 * @param linesize source plane linesize
209 * @param offx x-offsetting ranging in [-e;e]
210 * @param offy y-offsetting ranging in [-e;e]
211 * @param w source width
212 * @param h source height
213 * @param e research padding edge
215 static void compute_ssd_integral_image(const NLMeansDSPContext *dsp,
216 uint32_t *ii, ptrdiff_t ii_linesize_32,
217 const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
220 // ii has a surrounding padding of thickness "e"
221 const int ii_w = w + e*2;
222 const int ii_h = h + e*2;
224 // we center the first source
228 // 2nd source is the frame with offsetting
229 const int s2x = e + offx;
230 const int s2y = e + offy;
232 // get the dimension of the overlapping rectangle where it is always safe
233 // to compare the 2 sources pixels
234 const int startx_safe = FFMAX(s1x, s2x);
235 const int starty_safe = FFMAX(s1y, s2y);
236 const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
237 const int endy_safe = FFMIN(s1y + h, s2y + h);
239 // deduce the safe area width and height
240 const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
241 const int safe_ph = endy_safe - starty_safe;
243 // adjusted end x position of the safe area after width of the safe area gets aligned
244 const int endx_safe = startx_safe + safe_pw;
246 // top part where only one of s1 and s2 is still readable, or none at all
247 compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
253 // fill the left column integral required to compute the central
255 compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
259 startx_safe, safe_ph);
261 // main and safe part of the integral
262 av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
263 av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
264 av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
265 av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
266 if (safe_pw && safe_ph)
267 dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
268 src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
269 src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
272 // right part of the integral
273 compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
274 endx_safe, starty_safe,
277 ii_w - endx_safe, safe_ph);
279 // bottom part where only one of s1 and s2 is still readable, or none at all
280 compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
284 ii_w, ii_h - endy_safe);
287 static int config_input(AVFilterLink *inlink)
289 AVFilterContext *ctx = inlink->dst;
290 NLMeansContext *s = ctx->priv;
291 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
292 const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
293 + FFMAX(s->patch_hsize, s->patch_hsize_uv);
295 s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
296 s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
297 s->nb_planes = av_pix_fmt_count_planes(inlink->format);
299 /* Allocate the integral image with extra edges of thickness "e"
301 * +_+-------------------------------+
302 * |0|0000000000000000000000000000000|
303 * +-x-------------------------------+
307 * |0| +-----------------------+ |
312 * |0| +-----------------------+ |
316 * +-+-------------------------------+
318 s->ii_w = inlink->w + e*2;
319 s->ii_h = inlink->h + e*2;
321 // align to 4 the linesize, "+1" is for the space of the left 0-column
322 s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
324 // "+1" is for the space of the top 0-line
325 s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
327 return AVERROR(ENOMEM);
329 // skip top 0-line and left 0-column
330 s->ii = s->ii_orig + s->ii_lz_32 + 1;
332 // allocate weighted average for every pixel
333 s->wa_linesize = inlink->w;
334 s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
336 return AVERROR(ENOMEM);
343 ptrdiff_t src_linesize;
346 const uint32_t *ii_start;
350 static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
353 NLMeansContext *s = ctx->priv;
354 const struct thread_data *td = arg;
355 const ptrdiff_t src_linesize = td->src_linesize;
356 const int process_h = td->endy - td->starty;
357 const int slice_start = (process_h * jobnr ) / nb_jobs;
358 const int slice_end = (process_h * (jobnr+1)) / nb_jobs;
359 const int starty = td->starty + slice_start;
360 const int endy = td->starty + slice_end;
362 const uint32_t *ii = td->ii_start + (starty - p - 1) * s->ii_lz_32 - p - 1;
363 const int dist_b = 2*p + 1;
364 const int dist_d = dist_b * s->ii_lz_32;
365 const int dist_e = dist_d + dist_b;
367 for (y = starty; y < endy; y++) {
368 const uint8_t *src = td->src + y*src_linesize;
369 struct weighted_avg *wa = s->wa + y*s->wa_linesize;
370 for (x = td->startx; x < td->endx; x++) {
372 * M is a discrete map where every entry contains the sum of all the entries
373 * in the rectangle from the top-left origin of M to its coordinate. In the
374 * following schema, "i" contains the sum of the whole map:
376 * M = +----------+-----------------+----+
380 * +----------+-----------------+----+
386 * +----------+-----------------+----+
389 * +----------+-----------------+----+
391 * The sum of the X box can be calculated with:
394 * See https://en.wikipedia.org/wiki/Summed_area_table
396 * The compute*_ssd functions compute the integral image M where every entry
397 * contains the sum of the squared difference of every corresponding pixels of
398 * two input planes of the same size as M.
400 const uint32_t a = ii[x];
401 const uint32_t b = ii[x + dist_b];
402 const uint32_t d = ii[x + dist_d];
403 const uint32_t e = ii[x + dist_e];
404 const uint32_t patch_diff_sq = e - d - b + a;
406 if (patch_diff_sq < s->max_meaningful_diff) {
407 const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
408 wa[x].total_weight += weight;
409 wa[x].sum += weight * src[x];
417 static void weight_averages(uint8_t *dst, ptrdiff_t dst_linesize,
418 const uint8_t *src, ptrdiff_t src_linesize,
419 struct weighted_avg *wa, ptrdiff_t wa_linesize,
424 for (y = 0; y < h; y++) {
425 for (x = 0; x < w; x++) {
426 // Also weight the centered pixel
427 wa[x].total_weight += 1.f;
428 wa[x].sum += 1.f * src[x];
429 dst[x] = av_clip_uint8(wa[x].sum / wa[x].total_weight);
437 static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
438 uint8_t *dst, ptrdiff_t dst_linesize,
439 const uint8_t *src, ptrdiff_t src_linesize)
442 NLMeansContext *s = ctx->priv;
443 /* patches center points cover the whole research window so the patches
444 * themselves overflow the research window */
446 /* focus an integral pointer on the centered image (s1) */
447 const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;
449 memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));
451 for (offy = -r; offy <= r; offy++) {
452 for (offx = -r; offx <= r; offx++) {
454 struct thread_data td = {
455 .src = src + offy*src_linesize + offx,
456 .src_linesize = src_linesize,
457 .startx = FFMAX(0, -offx),
458 .starty = FFMAX(0, -offy),
459 .endx = FFMIN(w, w - offx),
460 .endy = FFMIN(h, h - offy),
461 .ii_start = centered_ii + offy*s->ii_lz_32 + offx,
465 compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32,
467 offx, offy, e, w, h);
468 ctx->internal->execute(ctx, nlmeans_slice, &td, NULL,
469 FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
474 weight_averages(dst, dst_linesize, src, src_linesize,
475 s->wa, s->wa_linesize, w, h);
480 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
483 AVFilterContext *ctx = inlink->dst;
484 NLMeansContext *s = ctx->priv;
485 AVFilterLink *outlink = ctx->outputs[0];
487 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
490 return AVERROR(ENOMEM);
492 av_frame_copy_props(out, in);
494 for (i = 0; i < s->nb_planes; i++) {
495 const int w = i ? s->chroma_w : inlink->w;
496 const int h = i ? s->chroma_h : inlink->h;
497 const int p = i ? s->patch_hsize_uv : s->patch_hsize;
498 const int r = i ? s->research_hsize_uv : s->research_hsize;
499 nlmeans_plane(ctx, w, h, p, r,
500 out->data[i], out->linesize[i],
501 in->data[i], in->linesize[i]);
505 return ff_filter_frame(outlink, out);
508 #define CHECK_ODD_FIELD(field, name) do { \
509 if (!(s->field & 1)) { \
511 av_log(ctx, AV_LOG_WARNING, name " size must be odd, " \
512 "setting it to %d\n", s->field); \
516 void ff_nlmeans_init(NLMeansDSPContext *dsp)
518 dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c;
521 ff_nlmeans_init_aarch64(dsp);
524 static av_cold int init(AVFilterContext *ctx)
527 NLMeansContext *s = ctx->priv;
528 const double h = s->sigma * 10.;
530 s->pdiff_scale = 1. / (h * h);
531 s->max_meaningful_diff = log(255.) / s->pdiff_scale;
532 av_assert0((s->max_meaningful_diff - 1) < FF_ARRAY_ELEMS(s->weight_lut));
533 for (i = 0; i < WEIGHT_LUT_SIZE; i++)
534 s->weight_lut[i] = exp(-i * s->pdiff_scale);
536 CHECK_ODD_FIELD(research_size, "Luma research window");
537 CHECK_ODD_FIELD(patch_size, "Luma patch");
539 if (!s->research_size_uv) s->research_size_uv = s->research_size;
540 if (!s->patch_size_uv) s->patch_size_uv = s->patch_size;
542 CHECK_ODD_FIELD(research_size_uv, "Chroma research window");
543 CHECK_ODD_FIELD(patch_size_uv, "Chroma patch");
545 s->research_hsize = s->research_size / 2;
546 s->research_hsize_uv = s->research_size_uv / 2;
547 s->patch_hsize = s->patch_size / 2;
548 s->patch_hsize_uv = s->patch_size_uv / 2;
550 av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
551 s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
552 s->patch_size, s->patch_size, s->patch_size_uv, s->patch_size_uv);
554 ff_nlmeans_init(&s->dsp);
559 static av_cold void uninit(AVFilterContext *ctx)
561 NLMeansContext *s = ctx->priv;
562 av_freep(&s->ii_orig);
566 static const AVFilterPad nlmeans_inputs[] = {
569 .type = AVMEDIA_TYPE_VIDEO,
570 .config_props = config_input,
571 .filter_frame = filter_frame,
576 static const AVFilterPad nlmeans_outputs[] = {
579 .type = AVMEDIA_TYPE_VIDEO,
584 AVFilter ff_vf_nlmeans = {
586 .description = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
587 .priv_size = sizeof(NLMeansContext),
590 .query_formats = query_formats,
591 .inputs = nlmeans_inputs,
592 .outputs = nlmeans_outputs,
593 .priv_class = &nlmeans_class,
594 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,