vp9_denoiser.c 12.6 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

Tim Kopp's avatar
Tim Kopp committed
18
19
20
21
#ifdef OUTPUT_YUV_DENOISED
static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
#endif

Tim Kopp's avatar
Tim Kopp committed
22
23
24
25
26
27
28
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};

int vp9_denoiser_filter() {
  return 0;
}

Tim Kopp's avatar
Tim Kopp committed
29
30
31
32
static int update_running_avg(const uint8_t *mc_avg, int mc_avg_stride,
                              uint8_t *avg, int avg_stride,
                              const uint8_t *sig, int sig_stride,
                              int increase_denoising, BLOCK_SIZE bs) {
Tim Kopp's avatar
Tim Kopp committed
33
  int r, c;
34
35
36
  int diff, adj, absdiff;
  int shift_inc1 = 0, shift_inc2 = 1;
  int adj_val[] = {3, 4, 6};
Tim Kopp's avatar
Tim Kopp committed
37
38
  int total_adj = 0;

39
40
41
42
43
  if (increase_denoising) {
    shift_inc1 = 1;
    shift_inc2 = 2;
  }

Tim Kopp's avatar
Tim Kopp committed
44
45
46
  for (r = 0; r < heights[bs]; ++r) {
    for (c = 0; c < widths[bs]; ++c) {
      diff = mc_avg[c] - sig[c];
47
      absdiff = abs(diff);
Tim Kopp's avatar
Tim Kopp committed
48

49
50
51
      if (absdiff <= 3 + shift_inc1) {
        avg[c] = mc_avg[c];
        total_adj += diff;
Tim Kopp's avatar
Tim Kopp committed
52
      } else {
53
54
55
        switch (absdiff) {
          case 4: case 5: case 6: case 7:
            adj = adj_val[0];
Tim Kopp's avatar
Tim Kopp committed
56
            break;
57
58
59
          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
60
            break;
61
62
63
64
65
66
67
68
69
70
          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
71
72
73
74
75
76
77
78
79
      }
    }
    sig += sig_stride;
    avg += avg_stride;
    mc_avg += mc_avg_stride;
  }
  return total_adj;
}

Tim Kopp's avatar
Tim Kopp committed
80
81
static uint8_t *block_start(uint8_t *framebuf, int stride,
                            int mi_row, int mi_col) {
82
  return framebuf + (stride * mi_row * 8) + (mi_col * 8);
83
84
85
86
87
88
89
90
91
92
93
94
95
96
}

void copy_block(uint8_t *dest, int dest_stride,
                uint8_t *src, int src_stride, BLOCK_SIZE bs) {
  int r, c;
  for (r = 0; r < heights[bs]; ++r) {
    for (c = 0; c < widths[bs]; ++c) {
      dest[c] = src[c];
    }
    dest += dest_stride;
    src += src_stride;
  }
}

97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
static int perform_motion_compensation(VP9_DENOISER *denoiser, MACROBLOCK *mb,
                                       BLOCK_SIZE bs, int increase_denoising,
                                       int mi_row, int mi_col) {
  // constants
  // TODO(tkopp): empirically determine good constants, or functions of block
  // size.
  int NOISE_MOTION_THRESHOLD = 25 * 25;
  int SSE_DIFF_THRESHOLD = heights[bs] * widths[bs] * 20;
  unsigned int SSE_THRESH = heights[bs] * widths[bs] * 40;
  unsigned int SSE_THRESH_HI = heights[bs] * widths[bs] * 60;

  int mv_col, mv_row;
  int sse_diff = denoiser->zero_mv_sse - denoiser->best_sse;
  int sse_diff_thresh;
  int sse_thresh;
  MV_REFERENCE_FRAME frame;
  MACROBLOCKD *filter_mbd = &mb->e_mbd;
  MB_MODE_INFO *mbmi = &filter_mbd->mi[0]->mbmi;

  // We will restore these after motion compensation.
  MB_MODE_INFO saved_mbmi = *mbmi;
  struct buf_2d saved_dst = filter_mbd->plane[0].dst;
  struct buf_2d saved_pre[2];
  saved_pre[0] = filter_mbd->plane[0].pre[0];
  saved_pre[1] = filter_mbd->plane[0].pre[1];

  // Decide the threshold for sum squared error.
  mv_col = denoiser->best_sse_mv.as_mv.col;
  mv_row = denoiser->best_sse_mv.as_mv.row;
  if (mv_row * mv_row + mv_col * mv_col > NOISE_MOTION_THRESHOLD) {
    sse_diff_thresh = 0;
  } else {
    sse_diff_thresh = SSE_DIFF_THRESHOLD;
  }

  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.
  if (frame != INTRA_FRAME && sse_diff > sse_diff_thresh) {
    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.
  filter_mbd->plane[0].pre[0].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[0].stride = denoiser->running_avg_y[frame].y_stride;

  filter_mbd->plane[1].pre[0].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[0].stride = denoiser->running_avg_y[frame].uv_stride;

  filter_mbd->plane[2].pre[0].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[0].stride = denoiser->running_avg_y[frame].uv_stride;

  filter_mbd->plane[0].pre[1].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[1].stride = denoiser->running_avg_y[frame].y_stride;

  filter_mbd->plane[1].pre[1].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[1].stride = denoiser->running_avg_y[frame].uv_stride;

  filter_mbd->plane[2].pre[1].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[1].stride = denoiser->running_avg_y[frame].uv_stride;

  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);
  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.y_stride;

  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);
  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.y_stride;

  vp9_build_inter_predictors_sby(filter_mbd, mv_row, mv_col, bs);

  // Restore everything to its original state
  filter_mbd->plane[0].pre[0] = saved_pre[0];
  filter_mbd->plane[0].pre[1] = saved_pre[1];
  filter_mbd->plane[0].dst = saved_dst;
  *mbmi = saved_mbmi;

  mv_row = denoiser->best_sse_mv.as_mv.row;
  mv_col = denoiser->best_sse_mv.as_mv.col;
  sse_thresh = denoiser->increase_denoising ? SSE_THRESH_HI : SSE_THRESH;

  // TODO(tkopp) why 8?
  if (denoiser->best_sse > sse_thresh ||
    mv_row * mv_row + mv_col * mv_col > 8 * NOISE_MOTION_THRESHOLD) {
    return COPY_BLOCK;
  }
  return FILTER_BLOCK;
}

229
void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
Tim Kopp's avatar
Tim Kopp committed
230
                          int mi_row, int mi_col, BLOCK_SIZE bs) {
231
232
233
  int decision = COPY_BLOCK;

  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
Tim Kopp's avatar
Tim Kopp committed
234
235
236
237
  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);
238
239
  struct buf_2d src = mb->plane[0].src;

240
241
242
  decision = perform_motion_compensation(denoiser, mb, bs,
                                         denoiser->increase_denoising,
                                         mi_row, mi_col);
Tim Kopp's avatar
Tim Kopp committed
243
  update_running_avg(mc_avg_start, mc_avg.y_stride, avg_start, avg.y_stride,
244
                     mb->plane[0].src.buf, mb->plane[0].src.stride, 0, bs);
245
246

  if (decision == FILTER_BLOCK) {
Tim Kopp's avatar
Tim Kopp committed
247
    // TODO(tkopp)
248
249
  }
  if (decision == COPY_BLOCK) {
Tim Kopp's avatar
Tim Kopp committed
250
    copy_block(avg_start, avg.y_stride, src.buf, src.stride, bs);
251
  }
Tim Kopp's avatar
Tim Kopp committed
252
253
}

Tim Kopp's avatar
Tim Kopp committed
254
static void copy_frame(YV12_BUFFER_CONFIG dest, const YV12_BUFFER_CONFIG src) {
255
  int r, c;
Tim Kopp's avatar
Tim Kopp committed
256
  const uint8_t *srcbuf = src.y_buffer;
257
258
259
260
261
262
263
264
265
266
267
  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) {
    for (c = 0; c < dest.y_width; ++c) {
      destbuf[c] = srcbuf[c];
    }
    destbuf += dest.y_stride;
    srcbuf += src.y_stride;
  }
268
269
}

Tim Kopp's avatar
Tim Kopp committed
270
void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
271
                                    YV12_BUFFER_CONFIG src,
Tim Kopp's avatar
Tim Kopp committed
272
273
274
275
                                    FRAME_TYPE frame_type,
                                    int refresh_alt_ref_frame,
                                    int refresh_golden_frame,
                                    int refresh_last_frame) {
276
  if (frame_type == KEY_FRAME) {
Tim Kopp's avatar
Tim Kopp committed
277
    int i;
278
279
280
281
282
    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
283
  } else {  /* For non key frames */
284
285
286
287
288
289
290
291
292
293
294
295
296
    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
297
298
}

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
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
318
319
320
}

int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
321
322
                       int ssx, int ssy, int border) {
  int i, fail;
Tim Kopp's avatar
Tim Kopp committed
323
  assert(denoiser != NULL);
324

325
326
  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
327
                                  ssx, ssy, border);
328
329
330
331
    if (fail) {
      vp9_denoiser_free(denoiser);
      return 1;
    }
Tim Kopp's avatar
Tim Kopp committed
332
333
334
#ifdef OUTPUT_YUV_DENOISED
    make_grayscale(&denoiser->running_avg_y[i]);
#endif
335
336
  }

337
  fail = vp9_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
Tim Kopp's avatar
Tim Kopp committed
338
                                ssx, ssy, border);
339
  if (fail) {
340
341
342
    vp9_denoiser_free(denoiser);
    return 1;
  }
Tim Kopp's avatar
Tim Kopp committed
343
344
345
#ifdef OUTPUT_YUV_DENOISED
  make_grayscale(&denoiser->running_avg_y[i]);
#endif
346
347
  denoiser->increase_denoising = 0;

Tim Kopp's avatar
Tim Kopp committed
348
349
350
351
  return 0;
}

void vp9_denoiser_free(VP9_DENOISER *denoiser) {
352
  int i;
Tim Kopp's avatar
Tim Kopp committed
353
354
355
  if (denoiser == NULL) {
    return;
  }
356
357
358
359
  for (i = 0; i < MAX_REF_FRAMES; ++i) {
    if (&denoiser->running_avg_y[i] != NULL) {
      vp9_free_frame_buffer(&denoiser->running_avg_y[i]);
    }
360
  }
361
362
  if (&denoiser->mc_running_avg_y != NULL) {
    vp9_free_frame_buffer(&denoiser->mc_running_avg_y);
363
  }
Tim Kopp's avatar
Tim Kopp committed
364
}
Tim Kopp's avatar
Tim Kopp committed
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383

#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