2 * Copyright (c) 2003-2013 Loren Merritt
3 * Copyright (c) 2015 Paul B Mahol
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (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 GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 /* Computes the Structural Similarity Metric between two video streams.
24 * Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli,
25 * "Image quality assessment: From error visibility to structural similarity,"
26 * IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, Apr. 2004.
28 * To improve speed, this implementation uses the standard approximation of
29 * overlapped 8x8 block sums, rather than the original gaussian weights.
34 * Caculate the SSIM between two input videos.
37 #include "libavutil/opt.h"
38 #include "libavutil/pixdesc.h"
40 #include "dualinput.h"
41 #include "drawutils.h"
46 typedef struct SSIMContext {
48 FFDualInputContext dinput;
62 #define OFFSET(x) offsetof(SSIMContext, x)
63 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
65 static const AVOption ssim_options[] = {
66 {"stats_file", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
67 {"f", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
71 AVFILTER_DEFINE_CLASS(ssim);
73 static int rgb_coefs[4] = { 1, 1, 1, 3};
74 static int yuv_coefs[4] = { 4, 1, 1, 6};
75 static int gray_coefs[4] = { 1, 0, 0, 1};
77 static void set_meta(AVDictionary **metadata, const char *key, char comp, float d)
80 snprintf(value, sizeof(value), "%0.2f", d);
83 snprintf(key2, sizeof(key2), "%s%c", key, comp);
84 av_dict_set(metadata, key2, value, 0);
86 av_dict_set(metadata, key, value, 0);
90 static void ssim_4x4x2_core(const uint8_t *main, int main_stride,
91 const uint8_t *ref, int ref_stride,
96 for (z = 0; z < 2; z++) {
97 uint32_t s1 = 0, s2 = 0, ss = 0, s12 = 0;
99 for (y = 0; y < 4; y++) {
100 for (x = 0; x < 4; x++) {
101 int a = main[x + y * main_stride];
102 int b = ref[x + y * ref_stride];
121 static float ssim_end1(int s1, int s2, int ss, int s12)
123 static const int ssim_c1 = (int)(.01*.01*255*255*64 + .5);
124 static const int ssim_c2 = (int)(.03*.03*255*255*64*63 + .5);
130 int vars = fss * 64 - fs1 * fs1 - fs2 * fs2;
131 int covar = fs12 * 64 - fs1 * fs2;
133 return (float)(2 * fs1 * fs2 + ssim_c1) * (float)(2 * covar + ssim_c2)
134 / ((float)(fs1 * fs1 + fs2 * fs2 + ssim_c1) * (float)(vars + ssim_c2));
137 static float ssim_end4(int sum0[5][4], int sum1[5][4], int width)
142 for( i = 0; i < width; i++ )
143 ssim += ssim_end1(sum0[i][0] + sum0[i + 1][0] + sum1[i][0] + sum1[i + 1][0],
144 sum0[i][1] + sum0[i + 1][1] + sum1[i][1] + sum1[i + 1][1],
145 sum0[i][2] + sum0[i + 1][2] + sum1[i][2] + sum1[i + 1][2],
146 sum0[i][3] + sum0[i + 1][3] + sum1[i][3] + sum1[i + 1][3]);
150 static float ssim_plane(uint8_t *main, int main_stride,
151 uint8_t *ref, int ref_stride,
152 int width, int height, void *temp)
157 int (*sum0)[4] = temp;
158 int (*sum1)[4] = sum0 + (width >> 2) + 3;
163 for (y = 1; y < height; y++) {
164 for (; z <= y; z++) {
165 FFSWAP(void*, sum0, sum1);
166 for (x = 0; x < width; x+=2)
167 ssim_4x4x2_core(&main[4 * (x + z * main_stride)], main_stride,
168 &ref[4 * (x + z * ref_stride)], ref_stride,
172 for (x = 0; x < width - 1; x += 4)
173 ssim += ssim_end4(sum0 + x, sum1 + x, FFMIN(4, width - x - 1));
176 return ssim / ((height - 1) * (width - 1));
179 static double ssim_db(double ssim, double weight)
181 return 10 * (log(weight) / log(10) - log(weight - ssim) / log(10));
184 static AVFrame *do_ssim(AVFilterContext *ctx, AVFrame *main,
187 AVDictionary **metadata = avpriv_frame_get_metadatap(main);
188 SSIMContext *s = ctx->priv;
194 for (i = 0; i < s->nb_components; i++)
195 c[i] = ssim_plane(main->data[i], main->linesize[i],
196 ref->data[i], ref->linesize[i],
197 s->planewidth[i], s->planeheight[i], s->temp);
199 ssimv = (c[0] * s->coefs[0] + c[1] * s->coefs[1] + c[2] * s->coefs[2]) / s->coefs[3];
201 for (i = 0; i < s->nb_components; i++)
202 set_meta(metadata, "lavfi.ssim.", s->comps[i], c[i]);
204 set_meta(metadata, "lavfi.ssim.All", 0, ssimv);
205 set_meta(metadata, "lavfi.ssim.dB", 0, ssim_db(c[0] * s->coefs[0] + c[1] * s->coefs[1] + c[2] * s->coefs[2], s->coefs[3]));
208 fprintf(s->stats_file, "n:%"PRId64" ", s->nb_frames);
210 for (i = 0; i < s->nb_components; i++)
211 fprintf(s->stats_file, "%c:%f ", s->comps[i], c[i]);
213 fprintf(s->stats_file, "All:%f (%f)\n", ssimv, ssim_db(c[0] * s->coefs[0] + c[1] * s->coefs[1] + c[2] * s->coefs[2], s->coefs[3]));
223 static av_cold int init(AVFilterContext *ctx)
225 SSIMContext *s = ctx->priv;
227 if (s->stats_file_str) {
228 s->stats_file = fopen(s->stats_file_str, "w");
229 if (!s->stats_file) {
230 int err = AVERROR(errno);
232 av_strerror(err, buf, sizeof(buf));
233 av_log(ctx, AV_LOG_ERROR, "Could not open stats file %s: %s\n",
234 s->stats_file_str, buf);
239 s->dinput.process = do_ssim;
240 s->dinput.shortest = 1;
241 s->dinput.repeatlast = 0;
245 static int query_formats(AVFilterContext *ctx)
247 static const enum AVPixelFormat pix_fmts[] = {
249 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
250 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV410P,
251 AV_PIX_FMT_YUVJ411P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
252 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
257 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
259 return AVERROR(ENOMEM);
260 return ff_set_common_formats(ctx, fmts_list);
263 static int config_input_ref(AVFilterLink *inlink)
265 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
266 AVFilterContext *ctx = inlink->dst;
267 SSIMContext *s = ctx->priv;
270 s->nb_components = desc->nb_components;
272 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
273 ctx->inputs[0]->h != ctx->inputs[1]->h) {
274 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
275 return AVERROR(EINVAL);
277 if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
278 av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
279 return AVERROR(EINVAL);
282 is_rgb = ff_fill_rgba_map(s->rgba_map, inlink->format) >= 0;
283 s->comps[0] = is_rgb ? 'R' : 'Y';
284 s->comps[1] = is_rgb ? 'G' : 'U';
285 s->comps[2] = is_rgb ? 'B' : 'V';
289 s->coefs = rgb_coefs;
290 } else if (s->nb_components == 1) {
291 s->coefs = gray_coefs;
293 s->coefs = yuv_coefs;
296 s->planeheight[1] = s->planeheight[2] = FF_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
297 s->planeheight[0] = s->planeheight[3] = inlink->h;
298 s->planewidth[1] = s->planewidth[2] = FF_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
299 s->planewidth[0] = s->planewidth[3] = inlink->w;
301 s->temp = av_malloc((2 * inlink->w + 12) * sizeof(*s->temp));
303 return AVERROR(ENOMEM);
308 static int config_output(AVFilterLink *outlink)
310 AVFilterContext *ctx = outlink->src;
311 SSIMContext *s = ctx->priv;
312 AVFilterLink *mainlink = ctx->inputs[0];
315 outlink->w = mainlink->w;
316 outlink->h = mainlink->h;
317 outlink->time_base = mainlink->time_base;
318 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
319 outlink->frame_rate = mainlink->frame_rate;
321 if ((ret = ff_dualinput_init(ctx, &s->dinput)) < 0)
327 static int filter_frame(AVFilterLink *inlink, AVFrame *buf)
329 SSIMContext *s = inlink->dst->priv;
330 return ff_dualinput_filter_frame(&s->dinput, inlink, buf);
333 static int request_frame(AVFilterLink *outlink)
335 SSIMContext *s = outlink->src->priv;
336 return ff_dualinput_request_frame(&s->dinput, outlink);
339 static av_cold void uninit(AVFilterContext *ctx)
341 SSIMContext *s = ctx->priv;
343 if (s->nb_frames > 0) {
344 if (s->nb_components == 3) {
345 av_log(ctx, AV_LOG_INFO, "SSIM %c:%f %c:%f %c:%f All:%f (%f)\n",
346 s->comps[0], s->ssim[0] / s->nb_frames,
347 s->comps[1], s->ssim[1] / s->nb_frames,
348 s->comps[2], s->ssim[2] / s->nb_frames,
349 (s->ssim[0] * 4 + s->ssim[1] + s->ssim[2]) / (s->nb_frames * 6),
350 ssim_db(s->ssim[0] * 4 + s->ssim[1] + s->ssim[2], s->nb_frames * 6));
351 } else if (s->nb_components == 1) {
352 av_log(ctx, AV_LOG_INFO, "SSIM All:%f (%f)\n",
353 s->ssim[0] / s->nb_frames, ssim_db(s->ssim[0], s->nb_frames));
357 ff_dualinput_uninit(&s->dinput);
360 fclose(s->stats_file);
365 static const AVFilterPad ssim_inputs[] = {
368 .type = AVMEDIA_TYPE_VIDEO,
369 .filter_frame = filter_frame,
372 .type = AVMEDIA_TYPE_VIDEO,
373 .filter_frame = filter_frame,
374 .config_props = config_input_ref,
379 static const AVFilterPad ssim_outputs[] = {
382 .type = AVMEDIA_TYPE_VIDEO,
383 .config_props = config_output,
384 .request_frame = request_frame,
389 AVFilter ff_vf_ssim = {
391 .description = NULL_IF_CONFIG_SMALL("Calculate the SSIM between two video streams."),
394 .query_formats = query_formats,
395 .priv_size = sizeof(SSIMContext),
396 .priv_class = &ssim_class,
397 .inputs = ssim_inputs,
398 .outputs = ssim_outputs,