blob: e0a953fba2cd9ae008b298f80d222134a6bc560f
1 | /* |
2 | * Copyright (c) 2007 Michael Niedermayer <michaelni@gmx.at> |
3 | * Copyright (c) 2013 Clément Bœsch <u pkh me> |
4 | * |
5 | * This file is part of FFmpeg. |
6 | * |
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. |
11 | * |
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. |
16 | * |
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. |
20 | */ |
21 | |
22 | /** |
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 |
29 | */ |
30 | |
31 | #include "libavutil/imgutils.h" |
32 | #include "libavutil/opt.h" |
33 | #include "libavutil/pixdesc.h" |
34 | #include "avfilter.h" |
35 | #include "internal.h" |
36 | |
37 | typedef struct { |
38 | const AVClass *class; |
39 | double luma_strength; |
40 | double chroma_strength; |
41 | int depth; |
42 | float *plane[16+1][4]; |
43 | int linesize; |
44 | int hsub, vsub; |
45 | int pixel_depth; |
46 | } OWDenoiseContext; |
47 | |
48 | #define OFFSET(x) offsetof(OWDenoiseContext, x) |
49 | #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM |
50 | static const AVOption owdenoise_options[] = { |
51 | { "depth", "set depth", OFFSET(depth), AV_OPT_TYPE_INT, {.i64 = 8}, 8, 16, FLAGS }, |
52 | { "luma_strength", "set luma strength", OFFSET(luma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS }, |
53 | { "ls", "set luma strength", OFFSET(luma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS }, |
54 | { "chroma_strength", "set chroma strength", OFFSET(chroma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS }, |
55 | { "cs", "set chroma strength", OFFSET(chroma_strength), AV_OPT_TYPE_DOUBLE, {.dbl = 1.0}, 0, 1000, FLAGS }, |
56 | { NULL } |
57 | }; |
58 | |
59 | AVFILTER_DEFINE_CLASS(owdenoise); |
60 | |
61 | DECLARE_ALIGNED(8, static const uint8_t, dither)[8][8] = { |
62 | { 0, 48, 12, 60, 3, 51, 15, 63 }, |
63 | { 32, 16, 44, 28, 35, 19, 47, 31 }, |
64 | { 8, 56, 4, 52, 11, 59, 7, 55 }, |
65 | { 40, 24, 36, 20, 43, 27, 39, 23 }, |
66 | { 2, 50, 14, 62, 1, 49, 13, 61 }, |
67 | { 34, 18, 46, 30, 33, 17, 45, 29 }, |
68 | { 10, 58, 6, 54, 9, 57, 5, 53 }, |
69 | { 42, 26, 38, 22, 41, 25, 37, 21 }, |
70 | }; |
71 | |
72 | static const double coeff[2][5] = { |
73 | { |
74 | 0.6029490182363579 * M_SQRT2, |
75 | 0.2668641184428723 * M_SQRT2, |
76 | -0.07822326652898785 * M_SQRT2, |
77 | -0.01686411844287495 * M_SQRT2, |
78 | 0.02674875741080976 * M_SQRT2, |
79 | },{ |
80 | 1.115087052456994 / M_SQRT2, |
81 | -0.5912717631142470 / M_SQRT2, |
82 | -0.05754352622849957 / M_SQRT2, |
83 | 0.09127176311424948 / M_SQRT2, |
84 | } |
85 | }; |
86 | |
87 | static const double icoeff[2][5] = { |
88 | { |
89 | 1.115087052456994 / M_SQRT2, |
90 | 0.5912717631142470 / M_SQRT2, |
91 | -0.05754352622849957 / M_SQRT2, |
92 | -0.09127176311424948 / M_SQRT2, |
93 | },{ |
94 | 0.6029490182363579 * M_SQRT2, |
95 | -0.2668641184428723 * M_SQRT2, |
96 | -0.07822326652898785 * M_SQRT2, |
97 | 0.01686411844287495 * M_SQRT2, |
98 | 0.02674875741080976 * M_SQRT2, |
99 | } |
100 | }; |
101 | |
102 | |
103 | static inline void decompose(float *dst_l, float *dst_h, const float *src, |
104 | int linesize, int w) |
105 | { |
106 | int x, i; |
107 | for (x = 0; x < w; x++) { |
108 | double sum_l = src[x * linesize] * coeff[0][0]; |
109 | double sum_h = src[x * linesize] * coeff[1][0]; |
110 | for (i = 1; i <= 4; i++) { |
111 | const double s = src[avpriv_mirror(x - i, w - 1) * linesize] |
112 | + src[avpriv_mirror(x + i, w - 1) * linesize]; |
113 | |
114 | sum_l += coeff[0][i] * s; |
115 | sum_h += coeff[1][i] * s; |
116 | } |
117 | dst_l[x * linesize] = sum_l; |
118 | dst_h[x * linesize] = sum_h; |
119 | } |
120 | } |
121 | |
122 | static inline void compose(float *dst, const float *src_l, const float *src_h, |
123 | int linesize, int w) |
124 | { |
125 | int x, i; |
126 | for (x = 0; x < w; x++) { |
127 | double sum_l = src_l[x * linesize] * icoeff[0][0]; |
128 | double sum_h = src_h[x * linesize] * icoeff[1][0]; |
129 | for (i = 1; i <= 4; i++) { |
130 | const int x0 = avpriv_mirror(x - i, w - 1) * linesize; |
131 | const int x1 = avpriv_mirror(x + i, w - 1) * linesize; |
132 | |
133 | sum_l += icoeff[0][i] * (src_l[x0] + src_l[x1]); |
134 | sum_h += icoeff[1][i] * (src_h[x0] + src_h[x1]); |
135 | } |
136 | dst[x * linesize] = (sum_l + sum_h) * 0.5; |
137 | } |
138 | } |
139 | |
140 | static inline void decompose2D(float *dst_l, float *dst_h, const float *src, |
141 | int xlinesize, int ylinesize, |
142 | int step, int w, int h) |
143 | { |
144 | int y, x; |
145 | for (y = 0; y < h; y++) |
146 | for (x = 0; x < step; x++) |
147 | decompose(dst_l + ylinesize*y + xlinesize*x, |
148 | dst_h + ylinesize*y + xlinesize*x, |
149 | src + ylinesize*y + xlinesize*x, |
150 | step * xlinesize, (w - x + step - 1) / step); |
151 | } |
152 | |
153 | static inline void compose2D(float *dst, const float *src_l, const float *src_h, |
154 | int xlinesize, int ylinesize, |
155 | int step, int w, int h) |
156 | { |
157 | int y, x; |
158 | for (y = 0; y < h; y++) |
159 | for (x = 0; x < step; x++) |
160 | compose(dst + ylinesize*y + xlinesize*x, |
161 | src_l + ylinesize*y + xlinesize*x, |
162 | src_h + ylinesize*y + xlinesize*x, |
163 | step * xlinesize, (w - x + step - 1) / step); |
164 | } |
165 | |
166 | static void decompose2D2(float *dst[4], float *src, float *temp[2], |
167 | int linesize, int step, int w, int h) |
168 | { |
169 | decompose2D(temp[0], temp[1], src, 1, linesize, step, w, h); |
170 | decompose2D( dst[0], dst[1], temp[0], linesize, 1, step, h, w); |
171 | decompose2D( dst[2], dst[3], temp[1], linesize, 1, step, h, w); |
172 | } |
173 | |
174 | static void compose2D2(float *dst, float *src[4], float *temp[2], |
175 | int linesize, int step, int w, int h) |
176 | { |
177 | compose2D(temp[0], src[0], src[1], linesize, 1, step, h, w); |
178 | compose2D(temp[1], src[2], src[3], linesize, 1, step, h, w); |
179 | compose2D(dst, temp[0], temp[1], 1, linesize, step, w, h); |
180 | } |
181 | |
182 | static void filter(OWDenoiseContext *s, |
183 | uint8_t *dst, int dst_linesize, |
184 | const uint8_t *src, int src_linesize, |
185 | int width, int height, double strength) |
186 | { |
187 | int x, y, i, j, depth = s->depth; |
188 | |
189 | while (1<<depth > width || 1<<depth > height) |
190 | depth--; |
191 | |
192 | if (s->pixel_depth <= 8) { |
193 | for (y = 0; y < height; y++) |
194 | for(x = 0; x < width; x++) |
195 | s->plane[0][0][y*s->linesize + x] = src[y*src_linesize + x]; |
196 | } else { |
197 | const uint16_t *src16 = (const uint16_t *)src; |
198 | |
199 | src_linesize /= 2; |
200 | for (y = 0; y < height; y++) |
201 | for(x = 0; x < width; x++) |
202 | s->plane[0][0][y*s->linesize + x] = src16[y*src_linesize + x]; |
203 | } |
204 | |
205 | for (i = 0; i < depth; i++) |
206 | decompose2D2(s->plane[i + 1], s->plane[i][0], s->plane[0] + 1, s->linesize, 1<<i, width, height); |
207 | |
208 | for (i = 0; i < depth; i++) { |
209 | for (j = 1; j < 4; j++) { |
210 | for (y = 0; y < height; y++) { |
211 | for (x = 0; x < width; x++) { |
212 | double v = s->plane[i + 1][j][y*s->linesize + x]; |
213 | if (v > strength) v -= strength; |
214 | else if (v < -strength) v += strength; |
215 | else v = 0; |
216 | s->plane[i + 1][j][x + y*s->linesize] = v; |
217 | } |
218 | } |
219 | } |
220 | } |
221 | for (i = depth-1; i >= 0; i--) |
222 | compose2D2(s->plane[i][0], s->plane[i + 1], s->plane[0] + 1, s->linesize, 1<<i, width, height); |
223 | |
224 | if (s->pixel_depth <= 8) { |
225 | for (y = 0; y < height; y++) { |
226 | for (x = 0; x < width; x++) { |
227 | 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 :) |
228 | if ((unsigned)i > 255U) i = ~(i >> 31); |
229 | dst[y*dst_linesize + x] = i; |
230 | } |
231 | } |
232 | } else { |
233 | uint16_t *dst16 = (uint16_t *)dst; |
234 | |
235 | dst_linesize /= 2; |
236 | for (y = 0; y < height; y++) { |
237 | for (x = 0; x < width; x++) { |
238 | i = s->plane[0][0][y*s->linesize + x]; |
239 | dst16[y*dst_linesize + x] = i; |
240 | } |
241 | } |
242 | } |
243 | } |
244 | |
245 | static int filter_frame(AVFilterLink *inlink, AVFrame *in) |
246 | { |
247 | AVFilterContext *ctx = inlink->dst; |
248 | OWDenoiseContext *s = ctx->priv; |
249 | AVFilterLink *outlink = ctx->outputs[0]; |
250 | AVFrame *out; |
251 | const int cw = AV_CEIL_RSHIFT(inlink->w, s->hsub); |
252 | const int ch = AV_CEIL_RSHIFT(inlink->h, s->vsub); |
253 | |
254 | if (av_frame_is_writable(in)) { |
255 | out = in; |
256 | |
257 | if (s->luma_strength > 0) |
258 | filter(s, out->data[0], out->linesize[0], in->data[0], in->linesize[0], inlink->w, inlink->h, s->luma_strength); |
259 | if (s->chroma_strength > 0) { |
260 | filter(s, out->data[1], out->linesize[1], in->data[1], in->linesize[1], cw, ch, s->chroma_strength); |
261 | filter(s, out->data[2], out->linesize[2], in->data[2], in->linesize[2], cw, ch, s->chroma_strength); |
262 | } |
263 | } else { |
264 | out = ff_get_video_buffer(outlink, outlink->w, outlink->h); |
265 | if (!out) { |
266 | av_frame_free(&in); |
267 | return AVERROR(ENOMEM); |
268 | } |
269 | av_frame_copy_props(out, in); |
270 | |
271 | if (s->luma_strength > 0) { |
272 | filter(s, out->data[0], out->linesize[0], in->data[0], in->linesize[0], inlink->w, inlink->h, s->luma_strength); |
273 | } else { |
274 | av_image_copy_plane(out->data[0], out->linesize[0], in ->data[0], in ->linesize[0], inlink->w, inlink->h); |
275 | } |
276 | if (s->chroma_strength > 0) { |
277 | filter(s, out->data[1], out->linesize[1], in->data[1], in->linesize[1], cw, ch, s->chroma_strength); |
278 | filter(s, out->data[2], out->linesize[2], in->data[2], in->linesize[2], cw, ch, s->chroma_strength); |
279 | } else { |
280 | av_image_copy_plane(out->data[1], out->linesize[1], in ->data[1], in ->linesize[1], inlink->w, inlink->h); |
281 | av_image_copy_plane(out->data[2], out->linesize[2], in ->data[2], in ->linesize[2], inlink->w, inlink->h); |
282 | } |
283 | |
284 | if (in->data[3]) |
285 | av_image_copy_plane(out->data[3], out->linesize[3], |
286 | in ->data[3], in ->linesize[3], |
287 | inlink->w, inlink->h); |
288 | av_frame_free(&in); |
289 | } |
290 | |
291 | return ff_filter_frame(outlink, out); |
292 | } |
293 | |
294 | static int query_formats(AVFilterContext *ctx) |
295 | { |
296 | static const enum AVPixelFormat pix_fmts[] = { |
297 | AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P, |
298 | AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV411P, |
299 | AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV440P, |
300 | AV_PIX_FMT_YUVA444P, AV_PIX_FMT_YUVA422P, |
301 | AV_PIX_FMT_YUVA420P, |
302 | AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9, |
303 | AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10, |
304 | AV_PIX_FMT_YUV440P10, |
305 | AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12, |
306 | AV_PIX_FMT_YUV440P12, |
307 | AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14, |
308 | AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16, |
309 | AV_PIX_FMT_NONE |
310 | }; |
311 | AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts); |
312 | if (!fmts_list) |
313 | return AVERROR(ENOMEM); |
314 | return ff_set_common_formats(ctx, fmts_list); |
315 | } |
316 | |
317 | static int config_input(AVFilterLink *inlink) |
318 | { |
319 | int i, j; |
320 | OWDenoiseContext *s = inlink->dst->priv; |
321 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format); |
322 | const int h = FFALIGN(inlink->h, 16); |
323 | |
324 | s->hsub = desc->log2_chroma_w; |
325 | s->vsub = desc->log2_chroma_h; |
326 | s->pixel_depth = desc->comp[0].depth; |
327 | |
328 | s->linesize = FFALIGN(inlink->w, 16); |
329 | for (j = 0; j < 4; j++) { |
330 | for (i = 0; i <= s->depth; i++) { |
331 | s->plane[i][j] = av_malloc_array(s->linesize, h * sizeof(s->plane[0][0][0])); |
332 | if (!s->plane[i][j]) |
333 | return AVERROR(ENOMEM); |
334 | } |
335 | } |
336 | return 0; |
337 | } |
338 | |
339 | static av_cold void uninit(AVFilterContext *ctx) |
340 | { |
341 | int i, j; |
342 | OWDenoiseContext *s = ctx->priv; |
343 | |
344 | for (j = 0; j < 4; j++) |
345 | for (i = 0; i <= s->depth; i++) |
346 | av_freep(&s->plane[i][j]); |
347 | } |
348 | |
349 | static const AVFilterPad owdenoise_inputs[] = { |
350 | { |
351 | .name = "default", |
352 | .type = AVMEDIA_TYPE_VIDEO, |
353 | .filter_frame = filter_frame, |
354 | .config_props = config_input, |
355 | }, |
356 | { NULL } |
357 | }; |
358 | |
359 | static const AVFilterPad owdenoise_outputs[] = { |
360 | { |
361 | .name = "default", |
362 | .type = AVMEDIA_TYPE_VIDEO, |
363 | }, |
364 | { NULL } |
365 | }; |
366 | |
367 | AVFilter ff_vf_owdenoise = { |
368 | .name = "owdenoise", |
369 | .description = NULL_IF_CONFIG_SMALL("Denoise using wavelets."), |
370 | .priv_size = sizeof(OWDenoiseContext), |
371 | .uninit = uninit, |
372 | .query_formats = query_formats, |
373 | .inputs = owdenoise_inputs, |
374 | .outputs = owdenoise_outputs, |
375 | .priv_class = &owdenoise_class, |
376 | .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC, |
377 | }; |
378 |