vp9_denoiser.c 14.7 KB
Newer Older
Tim Kopp's avatar
Tim Kopp committed
1
2
3
4
5
6
7
8
9
10
/*
 *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

Tim Kopp's avatar
Tim Kopp committed
11
#include <assert.h>
12
#include <limits.h>
Tim Kopp's avatar
Tim Kopp committed
13
#include "vpx_scale/yv12config.h"
Tim Kopp's avatar
Tim Kopp committed
14
#include "vpx/vpx_integer.h"
15
#include "vp9/common/vp9_reconinter.h"
Tim Kopp's avatar
Tim Kopp committed
16
#include "vp9/encoder/vp9_denoiser.h"
Tim Kopp's avatar
Tim Kopp committed
17

18
19
20
21
22
23
24
25
26
27
28
/* The VP9 denoiser is a work-in-progress. It currently is only designed to work
 * with speed 6, though it (inexplicably) seems to also work with speed 5 (one
 * would need to modify the source code in vp9_pickmode.c and vp9_encoder.c to
 * make the calls to the vp9_denoiser_* functions when in speed 5).
 *
 * The implementation is very similar to that of the VP8 denoiser. While
 * choosing the motion vectors / reference frames, the denoiser is run, and if
 * it did not modify the signal to much, the denoised block is copied to the
 * signal.
 */

Tim Kopp's avatar
Tim Kopp committed
29
30
31
32
#ifdef OUTPUT_YUV_DENOISED
static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
#endif

Tim Kopp's avatar
Tim Kopp committed
33
34
35
static const int widths[]  = {4, 4, 8, 8,  8, 16, 16, 16, 32, 32, 32, 64, 64};
static const int heights[] = {4, 8, 4, 8, 16,  8, 16, 32, 16, 32, 64, 32, 64};

36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) {
  (void)bs;
  return 3 + (increase_denoising ? 1 : 0);
}

static int delta_thresh(BLOCK_SIZE bs, int increase_denoising) {
  (void)bs;
  (void)increase_denoising;
  return 4;
}

static int noise_motion_thresh(BLOCK_SIZE bs, int increase_denoising) {
  (void)bs;
  (void)increase_denoising;
  return 25 * 25;
}

static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
  return widths[bs] * heights[bs] * (increase_denoising ? 60 : 40);
}

static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
                           int mv_row, int mv_col) {
  if (mv_row * mv_row + mv_col * mv_col >
      noise_motion_thresh(bs, increase_denoising)) {
    return 0;
  } else {
    return widths[bs] * heights[bs] * 20;
  }
}

static int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
}

static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
}

static VP9_DENOISER_DECISION denoiser_filter(const uint8_t *sig, int sig_stride,
                                             const uint8_t *mc_avg,
                                             int mc_avg_stride,
                                             uint8_t *avg, int avg_stride,
                                             int increase_denoising,
                                             BLOCK_SIZE bs) {
Tim Kopp's avatar
Tim Kopp committed
81
  int r, c;
82
83
84
85
  const uint8_t *sig_start = sig;
  const uint8_t *mc_avg_start = mc_avg;
  uint8_t *avg_start = avg;
  int diff, adj, absdiff, delta;
86
  int adj_val[] = {3, 4, 6};
Tim Kopp's avatar
Tim Kopp committed
87
88
  int total_adj = 0;

89
  // First attempt to apply a strong temporal denoising filter.
Tim Kopp's avatar
Tim Kopp committed
90
91
92
  for (r = 0; r < heights[bs]; ++r) {
    for (c = 0; c < widths[bs]; ++c) {
      diff = mc_avg[c] - sig[c];
93
      absdiff = abs(diff);
Tim Kopp's avatar
Tim Kopp committed
94

95
      if (absdiff <= absdiff_thresh(bs, increase_denoising)) {
96
97
        avg[c] = mc_avg[c];
        total_adj += diff;
Tim Kopp's avatar
Tim Kopp committed
98
      } else {
99
100
101
        switch (absdiff) {
          case 4: case 5: case 6: case 7:
            adj = adj_val[0];
Tim Kopp's avatar
Tim Kopp committed
102
            break;
103
104
105
          case 8: case 9: case 10: case 11:
          case 12: case 13: case 14: case 15:
            adj = adj_val[1];
Tim Kopp's avatar
Tim Kopp committed
106
            break;
107
108
109
110
111
112
113
114
115
116
          default:
            adj = adj_val[2];
        }
        if (diff > 0) {
          avg[c] = MIN(UINT8_MAX, sig[c] + adj);
          total_adj += adj;
        } else {
          avg[c] = MAX(0, sig[c] - adj);
          total_adj -= adj;
        }
Tim Kopp's avatar
Tim Kopp committed
117
118
119
120
121
122
      }
    }
    sig += sig_stride;
    avg += avg_stride;
    mc_avg += mc_avg_stride;
  }
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

  // If the strong filter did not modify the signal too much, we're all set.
  if (abs(total_adj) <= total_adj_strong_thresh(bs, increase_denoising)) {
    return FILTER_BLOCK;
  }

  // Otherwise, we try to dampen the filter if the delta is not too high.
  delta = ((abs(total_adj) - total_adj_strong_thresh(bs, increase_denoising))
           >> 8) + 1;
  if (delta > delta_thresh(bs, increase_denoising)) {
    return COPY_BLOCK;
  }

  mc_avg =  mc_avg_start;
  avg = avg_start;
  sig = sig_start;
  for (r = 0; r < heights[bs]; ++r) {
    for (c = 0; c < widths[bs]; ++c) {
      diff = mc_avg[c] - sig[c];
      adj = abs(diff);
      if (adj > delta) {
        adj = delta;
      }
      if (diff > 0) {
        avg[c] = MAX(0, avg[c] - adj);
        total_adj += adj;
      } else {
        avg[c] = MIN(UINT8_MAX, avg[c] + adj);
        total_adj -= adj;
      }
    }
    sig += sig_stride;
    avg += avg_stride;
    mc_avg += mc_avg_stride;
  }

  // We can use the filter if it has been sufficiently dampened
  if (abs(total_adj) <= total_adj_weak_thresh(bs, increase_denoising)) {
    return FILTER_BLOCK;
  }
  return COPY_BLOCK;
Tim Kopp's avatar
Tim Kopp committed
164
165
}

Tim Kopp's avatar
Tim Kopp committed
166
167
static uint8_t *block_start(uint8_t *framebuf, int stride,
                            int mi_row, int mi_col) {
168
  return framebuf + (stride * mi_row * 8) + (mi_col * 8);
169
170
}

Tim Kopp's avatar
Tim Kopp committed
171
172
173
static void copy_block(uint8_t *dest, int dest_stride,
                       const uint8_t *src, int src_stride, BLOCK_SIZE bs) {
  int r;
174
  for (r = 0; r < heights[bs]; ++r) {
Tim Kopp's avatar
Tim Kopp committed
175
    vpx_memcpy(dest, src, widths[bs]);
176
177
178
179
180
    dest += dest_stride;
    src += src_stride;
  }
}

181
182
183
184
185
186
static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
                                                         MACROBLOCK *mb,
                                                         BLOCK_SIZE bs,
                                                         int increase_denoising,
                                                         int mi_row,
                                                         int mi_col) {
187
188
189
190
191
192
  int mv_col, mv_row;
  int sse_diff = denoiser->zero_mv_sse - denoiser->best_sse;
  MV_REFERENCE_FRAME frame;
  MACROBLOCKD *filter_mbd = &mb->e_mbd;
  MB_MODE_INFO *mbmi = &filter_mbd->mi[0]->mbmi;

Tim Kopp's avatar
Tim Kopp committed
193
194
195
196
197
  MB_MODE_INFO saved_mbmi;
  int i, j;
  struct buf_2d saved_dst[MAX_MB_PLANE];
  struct buf_2d saved_pre[MAX_MB_PLANE][2];  // 2 pre buffers

198
  // We will restore these after motion compensation.
Tim Kopp's avatar
Tim Kopp committed
199
200
201
202
203
204
205
  saved_mbmi = *mbmi;
  for (i = 0; i < MAX_MB_PLANE; ++i) {
    for (j = 0; j < 2; ++j) {
      saved_pre[i][j] = filter_mbd->plane[i].pre[j];
    }
    saved_dst[i] = filter_mbd->plane[i].dst;
  }
206
207
208
209
210
211
212
213

  mv_col = denoiser->best_sse_mv.as_mv.col;
  mv_row = denoiser->best_sse_mv.as_mv.row;

  frame = denoiser->best_reference_frame;

  // If the best reference frame uses inter-prediction and there is enough of a
  // difference in sum-squared-error, use it.
214
215
  if (frame != INTRA_FRAME &&
      sse_diff > sse_diff_thresh(bs, increase_denoising, mv_row, mv_col)) {
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
    mbmi->ref_frame[0] = denoiser->best_reference_frame;
    mbmi->mode = denoiser->best_sse_inter_mode;
    mbmi->mv[0] = denoiser->best_sse_mv;
  } else {
    // Otherwise, use the zero reference frame.
    frame = denoiser->best_zeromv_reference_frame;

    mbmi->ref_frame[0] = denoiser->best_zeromv_reference_frame;
    mbmi->mode = ZEROMV;
    mbmi->mv[0].as_int = 0;

    denoiser->best_sse_inter_mode = ZEROMV;
    denoiser->best_sse_mv.as_int = 0;
    denoiser->best_sse = denoiser->zero_mv_sse;
  }

  // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
  // struct.
Tim Kopp's avatar
Tim Kopp committed
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
  for (j = 0; j < 2; ++j) {
    filter_mbd->plane[0].pre[j].buf =
        block_start(denoiser->running_avg_y[frame].y_buffer,
                    denoiser->running_avg_y[frame].y_stride,
                    mi_row, mi_col);
    filter_mbd->plane[0].pre[j].stride =
        denoiser->running_avg_y[frame].y_stride;
    filter_mbd->plane[1].pre[j].buf =
        block_start(denoiser->running_avg_y[frame].u_buffer,
                    denoiser->running_avg_y[frame].uv_stride,
                    mi_row, mi_col);
    filter_mbd->plane[1].pre[j].stride =
        denoiser->running_avg_y[frame].uv_stride;
    filter_mbd->plane[2].pre[j].buf =
        block_start(denoiser->running_avg_y[frame].v_buffer,
                    denoiser->running_avg_y[frame].uv_stride,
                    mi_row, mi_col);
    filter_mbd->plane[2].pre[j].stride =
        denoiser->running_avg_y[frame].uv_stride;
  }
254
255
256
257
258
259
260
261
262
  filter_mbd->plane[0].dst.buf =
      block_start(denoiser->mc_running_avg_y.y_buffer,
                  denoiser->mc_running_avg_y.y_stride,
                  mi_row, mi_col);
  filter_mbd->plane[0].dst.stride = denoiser->mc_running_avg_y.y_stride;
  filter_mbd->plane[1].dst.buf =
      block_start(denoiser->mc_running_avg_y.u_buffer,
                  denoiser->mc_running_avg_y.uv_stride,
                  mi_row, mi_col);
Tim Kopp's avatar
Tim Kopp committed
263
  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.uv_stride;
264
265
266
267
  filter_mbd->plane[2].dst.buf =
      block_start(denoiser->mc_running_avg_y.v_buffer,
                  denoiser->mc_running_avg_y.uv_stride,
                  mi_row, mi_col);
Tim Kopp's avatar
Tim Kopp committed
268
  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.uv_stride;
269
270
271
272
273

  vp9_build_inter_predictors_sby(filter_mbd, mv_row, mv_col, bs);

  // Restore everything to its original state
  *mbmi = saved_mbmi;
Tim Kopp's avatar
Tim Kopp committed
274
275
276
277
278
279
  for (i = 0; i < MAX_MB_PLANE; ++i) {
    for (j = 0; j < 2; ++j) {
      filter_mbd->plane[i].pre[j] = saved_pre[i][j];
    }
    filter_mbd->plane[i].dst = saved_dst[i];
  }
280
281
282
283

  mv_row = denoiser->best_sse_mv.as_mv.row;
  mv_col = denoiser->best_sse_mv.as_mv.col;

284
285
286
287
288
  if (denoiser->best_sse > sse_thresh(bs, increase_denoising)) {
    return COPY_BLOCK;
  }
  if (mv_row * mv_row + mv_col * mv_col >
      8 * noise_motion_thresh(bs, increase_denoising)) {
289
290
291
292
293
    return COPY_BLOCK;
  }
  return FILTER_BLOCK;
}

294
void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
Tim Kopp's avatar
Tim Kopp committed
295
                          int mi_row, int mi_col, BLOCK_SIZE bs) {
296
  VP9_DENOISER_DECISION decision = FILTER_BLOCK;
297
  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
Tim Kopp's avatar
Tim Kopp committed
298
299
300
301
  YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y;
  uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col);
  uint8_t *mc_avg_start = block_start(mc_avg.y_buffer, mc_avg.y_stride,
                                          mi_row, mi_col);
302
303
  struct buf_2d src = mb->plane[0].src;

304
305
306
  decision = perform_motion_compensation(denoiser, mb, bs,
                                         denoiser->increase_denoising,
                                         mi_row, mi_col);
307
308

  if (decision == FILTER_BLOCK) {
309
310
311
312
    decision = denoiser_filter(src.buf, src.stride,
                               mc_avg_start, mc_avg.y_stride,
                               avg_start, avg.y_stride,
                               0, bs);
313
  }
314
315
316
317

  if (decision == FILTER_BLOCK) {
    copy_block(src.buf, src.stride, avg_start, avg.y_stride, bs);
  } else {  // COPY_BLOCK
Tim Kopp's avatar
Tim Kopp committed
318
    copy_block(avg_start, avg.y_stride, src.buf, src.stride, bs);
319
  }
Tim Kopp's avatar
Tim Kopp committed
320
321
}

Tim Kopp's avatar
Tim Kopp committed
322
static void copy_frame(YV12_BUFFER_CONFIG dest, const YV12_BUFFER_CONFIG src) {
Tim Kopp's avatar
Tim Kopp committed
323
  int r;
Tim Kopp's avatar
Tim Kopp committed
324
  const uint8_t *srcbuf = src.y_buffer;
325
326
327
328
329
  uint8_t *destbuf = dest.y_buffer;
  assert(dest.y_width == src.y_width);
  assert(dest.y_height == src.y_height);

  for (r = 0; r < dest.y_height; ++r) {
Tim Kopp's avatar
Tim Kopp committed
330
    vpx_memcpy(destbuf, srcbuf, dest.y_width);
331
332
333
    destbuf += dest.y_stride;
    srcbuf += src.y_stride;
  }
334
335
}

Tim Kopp's avatar
Tim Kopp committed
336
void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
337
                                    YV12_BUFFER_CONFIG src,
Tim Kopp's avatar
Tim Kopp committed
338
339
340
341
                                    FRAME_TYPE frame_type,
                                    int refresh_alt_ref_frame,
                                    int refresh_golden_frame,
                                    int refresh_last_frame) {
342
  if (frame_type == KEY_FRAME) {
Tim Kopp's avatar
Tim Kopp committed
343
    int i;
344
345
346
347
348
    copy_frame(denoiser->running_avg_y[LAST_FRAME], src);
    for (i = 2; i < MAX_REF_FRAMES - 1; i++) {
      copy_frame(denoiser->running_avg_y[i],
                 denoiser->running_avg_y[LAST_FRAME]);
    }
Tim Kopp's avatar
Tim Kopp committed
349
  } else {  /* For non key frames */
350
351
352
353
354
355
356
357
358
359
360
361
362
    if (refresh_alt_ref_frame) {
      copy_frame(denoiser->running_avg_y[ALTREF_FRAME],
                 denoiser->running_avg_y[INTRA_FRAME]);
    }
    if (refresh_golden_frame) {
      copy_frame(denoiser->running_avg_y[GOLDEN_FRAME],
                 denoiser->running_avg_y[INTRA_FRAME]);
    }
    if (refresh_last_frame) {
      copy_frame(denoiser->running_avg_y[LAST_FRAME],
                 denoiser->running_avg_y[INTRA_FRAME]);
    }
  }
Tim Kopp's avatar
Tim Kopp committed
363
364
}

365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
void vp9_denoiser_reset_frame_stats(VP9_DENOISER *denoiser) {
  denoiser->zero_mv_sse = UINT_MAX;
  denoiser->best_sse = UINT_MAX;
}

void vp9_denoiser_update_frame_stats(VP9_DENOISER *denoiser, MB_MODE_INFO *mbmi,
                                     unsigned int sse, PREDICTION_MODE mode) {
  // TODO(tkopp): Use both MVs if possible
  if (mbmi->mv[0].as_int == 0 && sse < denoiser->zero_mv_sse) {
    denoiser->zero_mv_sse = sse;
    denoiser->best_zeromv_reference_frame = mbmi->ref_frame[0];
  }

  if (mbmi->mv[0].as_int != 0 && sse < denoiser->best_sse) {
    denoiser->best_sse = sse;
    denoiser->best_sse_inter_mode = mode;
    denoiser->best_sse_mv = mbmi->mv[0];
    denoiser->best_reference_frame = mbmi->ref_frame[0];
  }
Tim Kopp's avatar
Tim Kopp committed
384
385
386
}

int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
387
388
                       int ssx, int ssy, int border) {
  int i, fail;
Tim Kopp's avatar
Tim Kopp committed
389
  assert(denoiser != NULL);
390

391
392
  for (i = 0; i < MAX_REF_FRAMES; ++i) {
    fail = vp9_alloc_frame_buffer(&denoiser->running_avg_y[i], width, height,
Tim Kopp's avatar
Tim Kopp committed
393
                                  ssx, ssy, border);
394
395
396
397
    if (fail) {
      vp9_denoiser_free(denoiser);
      return 1;
    }
Tim Kopp's avatar
Tim Kopp committed
398
399
400
#ifdef OUTPUT_YUV_DENOISED
    make_grayscale(&denoiser->running_avg_y[i]);
#endif
401
402
  }

403
  fail = vp9_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
Tim Kopp's avatar
Tim Kopp committed
404
                                ssx, ssy, border);
405
  if (fail) {
406
407
408
    vp9_denoiser_free(denoiser);
    return 1;
  }
Tim Kopp's avatar
Tim Kopp committed
409
410
411
#ifdef OUTPUT_YUV_DENOISED
  make_grayscale(&denoiser->running_avg_y[i]);
#endif
412
413
  denoiser->increase_denoising = 0;

Tim Kopp's avatar
Tim Kopp committed
414
415
416
417
  return 0;
}

void vp9_denoiser_free(VP9_DENOISER *denoiser) {
418
  int i;
Tim Kopp's avatar
Tim Kopp committed
419
420
421
  if (denoiser == NULL) {
    return;
  }
422
423
424
425
  for (i = 0; i < MAX_REF_FRAMES; ++i) {
    if (&denoiser->running_avg_y[i] != NULL) {
      vp9_free_frame_buffer(&denoiser->running_avg_y[i]);
    }
426
  }
427
428
  if (&denoiser->mc_running_avg_y != NULL) {
    vp9_free_frame_buffer(&denoiser->mc_running_avg_y);
429
  }
Tim Kopp's avatar
Tim Kopp committed
430
}
Tim Kopp's avatar
Tim Kopp committed
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449

#ifdef OUTPUT_YUV_DENOISED
static void make_grayscale(YV12_BUFFER_CONFIG *yuv) {
  int r, c;
  uint8_t *u = yuv->u_buffer;
  uint8_t *v = yuv->v_buffer;

  // The '/2's are there because we have a 440 buffer, but we want to output
  // 420.
  for (r = 0; r < yuv->uv_height / 2; ++r) {
    for (c = 0; c < yuv->uv_width / 2; ++c) {
      u[c] = UINT8_MAX / 2;
      v[c] = UINT8_MAX / 2;
    }
    u += yuv->uv_stride + yuv->uv_width / 2;
    v += yuv->uv_stride + yuv->uv_width / 2;
  }
}
#endif