2 * Copyright (c) 2007 Michael Niedermayer <michaelni@gmx.at>
3 * Copyright (c) 2013 Clément Bœsch <u pkh me>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (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
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License along
18 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 * @todo try to change to int
24 * @todo try lifting based implementation
25 * @todo optimize optimize optimize
26 * @todo hard thresholding
27 * @todo use QP to decide filter strength
28 * @todo wavelet normalization / least squares optimal signal vs. noise thresholds
31 #include "libavutil/imgutils.h"
32 #include "libavutil/opt.h"
33 #include "libavutil/mem_internal.h"
34 #include "libavutil/pixdesc.h"
38 typedef struct OWDenoiseContext {
41 double chroma_strength;
43 float *plane[16+1][4];
49 #define OFFSET(x) offsetof(OWDenoiseContext, x)
50 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
51 static const AVOption owdenoise_options[] = {
52 { "depth", "set depth", OFFSET(depth), AV_OPT_TYPE_INT, {.i64 = 8}, 8, 16, FLAGS },
53 { "luma_strength", "set luma strength", OFFSET(luma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS },
54 { "ls", "set luma strength", OFFSET(luma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS },
55 { "chroma_strength", "set chroma strength", OFFSET(chroma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS },
56 { "cs", "set chroma strength", OFFSET(chroma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS },
60 AVFILTER_DEFINE_CLASS(owdenoise);
62 DECLARE_ALIGNED(8, static const uint8_t, dither)[8][8] = {
63 { 0, 48, 12, 60, 3, 51, 15, 63 },
64 { 32, 16, 44, 28, 35, 19, 47, 31 },
65 { 8, 56, 4, 52, 11, 59, 7, 55 },
66 { 40, 24, 36, 20, 43, 27, 39, 23 },
67 { 2, 50, 14, 62, 1, 49, 13, 61 },
68 { 34, 18, 46, 30, 33, 17, 45, 29 },
69 { 10, 58, 6, 54, 9, 57, 5, 53 },
70 { 42, 26, 38, 22, 41, 25, 37, 21 },
73 static const double coeff[2][5] = {
75 0.6029490182363579 * M_SQRT2,
76 0.2668641184428723 * M_SQRT2,
77 -0.07822326652898785 * M_SQRT2,
78 -0.01686411844287495 * M_SQRT2,
79 0.02674875741080976 * M_SQRT2,
81 1.115087052456994 / M_SQRT2,
82 -0.5912717631142470 / M_SQRT2,
83 -0.05754352622849957 / M_SQRT2,
84 0.09127176311424948 / M_SQRT2,
88 static const double icoeff[2][5] = {
90 1.115087052456994 / M_SQRT2,
91 0.5912717631142470 / M_SQRT2,
92 -0.05754352622849957 / M_SQRT2,
93 -0.09127176311424948 / M_SQRT2,
95 0.6029490182363579 * M_SQRT2,
96 -0.2668641184428723 * M_SQRT2,
97 -0.07822326652898785 * M_SQRT2,
98 0.01686411844287495 * M_SQRT2,
99 0.02674875741080976 * M_SQRT2,
104 static inline void decompose(float *dst_l, float *dst_h, const float *src,
108 for (x = 0; x < w; x++) {
109 double sum_l = src[x * linesize] * coeff[0][0];
110 double sum_h = src[x * linesize] * coeff[1][0];
111 for (i = 1; i <= 4; i++) {
112 const double s = src[avpriv_mirror(x - i, w - 1) * linesize]
113 + src[avpriv_mirror(x + i, w - 1) * linesize];
115 sum_l += coeff[0][i] * s;
116 sum_h += coeff[1][i] * s;
118 dst_l[x * linesize] = sum_l;
119 dst_h[x * linesize] = sum_h;
123 static inline void compose(float *dst, const float *src_l, const float *src_h,
127 for (x = 0; x < w; x++) {
128 double sum_l = src_l[x * linesize] * icoeff[0][0];
129 double sum_h = src_h[x * linesize] * icoeff[1][0];
130 for (i = 1; i <= 4; i++) {
131 const int x0 = avpriv_mirror(x - i, w - 1) * linesize;
132 const int x1 = avpriv_mirror(x + i, w - 1) * linesize;
134 sum_l += icoeff[0][i] * (src_l[x0] + src_l[x1]);
135 sum_h += icoeff[1][i] * (src_h[x0] + src_h[x1]);
137 dst[x * linesize] = (sum_l + sum_h) * 0.5;
141 static inline void decompose2D(float *dst_l, float *dst_h, const float *src,
142 int xlinesize, int ylinesize,
143 int step, int w, int h)
146 for (y = 0; y < h; y++)
147 for (x = 0; x < step; x++)
148 decompose(dst_l + ylinesize*y + xlinesize*x,
149 dst_h + ylinesize*y + xlinesize*x,
150 src + ylinesize*y + xlinesize*x,
151 step * xlinesize, (w - x + step - 1) / step);
154 static inline void compose2D(float *dst, const float *src_l, const float *src_h,
155 int xlinesize, int ylinesize,
156 int step, int w, int h)
159 for (y = 0; y < h; y++)
160 for (x = 0; x < step; x++)
161 compose(dst + ylinesize*y + xlinesize*x,
162 src_l + ylinesize*y + xlinesize*x,
163 src_h + ylinesize*y + xlinesize*x,
164 step * xlinesize, (w - x + step - 1) / step);
167 static void decompose2D2(float *dst[4], float *src, float *temp[2],
168 int linesize, int step, int w, int h)
170 decompose2D(temp[0], temp[1], src, 1, linesize, step, w, h);
171 decompose2D( dst[0], dst[1], temp[0], linesize, 1, step, h, w);
172 decompose2D( dst[2], dst[3], temp[1], linesize, 1, step, h, w);
175 static void compose2D2(float *dst, float *src[4], float *temp[2],
176 int linesize, int step, int w, int h)
178 compose2D(temp[0], src[0], src[1], linesize, 1, step, h, w);
179 compose2D(temp[1], src[2], src[3], linesize, 1, step, h, w);
180 compose2D(dst, temp[0], temp[1], 1, linesize, step, w, h);
183 static void filter(OWDenoiseContext *s,
184 uint8_t *dst, int dst_linesize,
185 const uint8_t *src, int src_linesize,
186 int width, int height, double strength)
188 int x, y, i, j, depth = s->depth;
190 while (1<<depth > width || 1<<depth > height)
193 if (s->pixel_depth <= 8) {
194 for (y = 0; y < height; y++)
195 for(x = 0; x < width; x++)
196 s->plane[0][0][y*s->linesize + x] = src[y*src_linesize + x];
198 const uint16_t *src16 = (const uint16_t *)src;
201 for (y = 0; y < height; y++)
202 for(x = 0; x < width; x++)
203 s->plane[0][0][y*s->linesize + x] = src16[y*src_linesize + x];
206 for (i = 0; i < depth; i++)
207 decompose2D2(s->plane[i + 1], s->plane[i][0], s->plane[0] + 1, s->linesize, 1<<i, width, height);
209 for (i = 0; i < depth; i++) {
210 for (j = 1; j < 4; j++) {
211 for (y = 0; y < height; y++) {
212 for (x = 0; x < width; x++) {
213 double v = s->plane[i + 1][j][y*s->linesize + x];
214 if (v > strength) v -= strength;
215 else if (v < -strength) v += strength;
217 s->plane[i + 1][j][x + y*s->linesize] = v;
222 for (i = depth-1; i >= 0; i--)
223 compose2D2(s->plane[i][0], s->plane[i + 1], s->plane[0] + 1, s->linesize, 1<<i, width, height);
225 if (s->pixel_depth <= 8) {
226 for (y = 0; y < height; y++) {
227 for (x = 0; x < width; x++) {
228 i = s->plane[0][0][y*s->linesize + x] + dither[x&7][y&7]*(1.0/64) + 1.0/128; // yes the rounding is insane but optimal :)
229 if ((unsigned)i > 255U) i = ~(i >> 31);
230 dst[y*dst_linesize + x] = i;
234 uint16_t *dst16 = (uint16_t *)dst;
237 for (y = 0; y < height; y++) {
238 for (x = 0; x < width; x++) {
239 i = s->plane[0][0][y*s->linesize + x];
240 dst16[y*dst_linesize + x] = i;
246 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
248 AVFilterContext *ctx = inlink->dst;
249 OWDenoiseContext *s = ctx->priv;
250 AVFilterLink *outlink = ctx->outputs[0];
252 const int cw = AV_CEIL_RSHIFT(inlink->w, s->hsub);
253 const int ch = AV_CEIL_RSHIFT(inlink->h, s->vsub);
255 if (av_frame_is_writable(in)) {
258 if (s->luma_strength > 0)
259 filter(s, out->data[0], out->linesize[0], in->data[0], in->linesize[0], inlink->w, inlink->h, s->luma_strength);
260 if (s->chroma_strength > 0) {
261 filter(s, out->data[1], out->linesize[1], in->data[1], in->linesize[1], cw, ch, s->chroma_strength);
262 filter(s, out->data[2], out->linesize[2], in->data[2], in->linesize[2], cw, ch, s->chroma_strength);
265 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
268 return AVERROR(ENOMEM);
270 av_frame_copy_props(out, in);
272 if (s->luma_strength > 0) {
273 filter(s, out->data[0], out->linesize[0], in->data[0], in->linesize[0], inlink->w, inlink->h, s->luma_strength);
275 av_image_copy_plane(out->data[0], out->linesize[0], in ->data[0], in ->linesize[0], inlink->w, inlink->h);
277 if (s->chroma_strength > 0) {
278 filter(s, out->data[1], out->linesize[1], in->data[1], in->linesize[1], cw, ch, s->chroma_strength);
279 filter(s, out->data[2], out->linesize[2], in->data[2], in->linesize[2], cw, ch, s->chroma_strength);
281 av_image_copy_plane(out->data[1], out->linesize[1], in ->data[1], in ->linesize[1], inlink->w, inlink->h);
282 av_image_copy_plane(out->data[2], out->linesize[2], in ->data[2], in ->linesize[2], inlink->w, inlink->h);
286 av_image_copy_plane(out->data[3], out->linesize[3],
287 in ->data[3], in ->linesize[3],
288 inlink->w, inlink->h);
292 return ff_filter_frame(outlink, out);
295 static int query_formats(AVFilterContext *ctx)
297 static const enum AVPixelFormat pix_fmts[] = {
298 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P,
299 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV411P,
300 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV440P,
301 AV_PIX_FMT_YUVA444P, AV_PIX_FMT_YUVA422P,
303 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
304 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
305 AV_PIX_FMT_YUV440P10,
306 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
307 AV_PIX_FMT_YUV440P12,
308 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
309 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
312 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
314 return AVERROR(ENOMEM);
315 return ff_set_common_formats(ctx, fmts_list);
318 static int config_input(AVFilterLink *inlink)
321 OWDenoiseContext *s = inlink->dst->priv;
322 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
323 const int h = FFALIGN(inlink->h, 16);
325 s->hsub = desc->log2_chroma_w;
326 s->vsub = desc->log2_chroma_h;
327 s->pixel_depth = desc->comp[0].depth;
329 s->linesize = FFALIGN(inlink->w, 16);
330 for (j = 0; j < 4; j++) {
331 for (i = 0; i <= s->depth; i++) {
332 s->plane[i][j] = av_malloc_array(s->linesize, h * sizeof(s->plane[0][0][0]));
334 return AVERROR(ENOMEM);
340 static av_cold void uninit(AVFilterContext *ctx)
343 OWDenoiseContext *s = ctx->priv;
345 for (j = 0; j < 4; j++)
346 for (i = 0; i <= s->depth; i++)
347 av_freep(&s->plane[i][j]);
350 static const AVFilterPad owdenoise_inputs[] = {
353 .type = AVMEDIA_TYPE_VIDEO,
354 .filter_frame = filter_frame,
355 .config_props = config_input,
360 static const AVFilterPad owdenoise_outputs[] = {
363 .type = AVMEDIA_TYPE_VIDEO,
368 const AVFilter ff_vf_owdenoise = {
370 .description = NULL_IF_CONFIG_SMALL("Denoise using wavelets."),
371 .priv_size = sizeof(OWDenoiseContext),
373 .query_formats = query_formats,
374 .inputs = owdenoise_inputs,
375 .outputs = owdenoise_outputs,
376 .priv_class = &owdenoise_class,
377 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,