1 /*
2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <assert.h>
12 #include <limits.h>
13 #include <math.h>
14
15 #include "./vpx_dsp_rtcd.h"
16 #include "vpx_dsp/vpx_dsp_common.h"
17 #include "vpx_scale/yv12config.h"
18 #include "vpx/vpx_integer.h"
19 #include "vp9/common/vp9_reconinter.h"
20 #include "vp9/encoder/vp9_context_tree.h"
21 #include "vp9/encoder/vp9_noise_estimate.h"
22 #include "vp9/encoder/vp9_encoder.h"
23
24 #if CONFIG_VP9_TEMPORAL_DENOISING
25 // For SVC: only do noise estimation on top spatial layer.
noise_est_svc(const struct VP9_COMP * const cpi)26 static INLINE int noise_est_svc(const struct VP9_COMP *const cpi) {
27 return (!cpi->use_svc ||
28 (cpi->use_svc &&
29 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1));
30 }
31 #endif
32
vp9_noise_estimate_init(NOISE_ESTIMATE * const ne,int width,int height)33 void vp9_noise_estimate_init(NOISE_ESTIMATE *const ne, int width, int height) {
34 ne->enabled = 0;
35 ne->level = (width * height < 1280 * 720) ? kLowLow : kLow;
36 ne->value = 0;
37 ne->count = 0;
38 ne->thresh = 90;
39 ne->last_w = 0;
40 ne->last_h = 0;
41 if (width * height >= 1920 * 1080) {
42 ne->thresh = 200;
43 } else if (width * height >= 1280 * 720) {
44 ne->thresh = 140;
45 } else if (width * height >= 640 * 360) {
46 ne->thresh = 115;
47 }
48 ne->num_frames_estimate = 15;
49 ne->adapt_thresh = (3 * ne->thresh) >> 1;
50 }
51
enable_noise_estimation(VP9_COMP * const cpi)52 static int enable_noise_estimation(VP9_COMP *const cpi) {
53 #if CONFIG_VP9_HIGHBITDEPTH
54 if (cpi->common.use_highbitdepth) return 0;
55 #endif
56 // Enable noise estimation if denoising is on.
57 #if CONFIG_VP9_TEMPORAL_DENOISING
58 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi) &&
59 cpi->common.width >= 320 && cpi->common.height >= 180)
60 return 1;
61 #endif
62 // Only allow noise estimate under certain encoding mode.
63 // Enabled for 1 pass CBR, speed >=5, and if resolution is same as original.
64 // Not enabled for SVC mode and screen_content_mode.
65 // Not enabled for low resolutions.
66 if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR &&
67 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cpi->oxcf.speed >= 5 &&
68 cpi->resize_state == ORIG && cpi->resize_pending == 0 && !cpi->use_svc &&
69 cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
70 cpi->common.width * cpi->common.height >= 640 * 360)
71 return 1;
72 else
73 return 0;
74 }
75
76 #if CONFIG_VP9_TEMPORAL_DENOISING
copy_frame(YV12_BUFFER_CONFIG * const dest,const YV12_BUFFER_CONFIG * const src)77 static void copy_frame(YV12_BUFFER_CONFIG *const dest,
78 const YV12_BUFFER_CONFIG *const src) {
79 int r;
80 const uint8_t *srcbuf = src->y_buffer;
81 uint8_t *destbuf = dest->y_buffer;
82
83 assert(dest->y_width == src->y_width);
84 assert(dest->y_height == src->y_height);
85
86 for (r = 0; r < dest->y_height; ++r) {
87 memcpy(destbuf, srcbuf, dest->y_width);
88 destbuf += dest->y_stride;
89 srcbuf += src->y_stride;
90 }
91 }
92 #endif // CONFIG_VP9_TEMPORAL_DENOISING
93
vp9_noise_estimate_extract_level(NOISE_ESTIMATE * const ne)94 NOISE_LEVEL vp9_noise_estimate_extract_level(NOISE_ESTIMATE *const ne) {
95 int noise_level = kLowLow;
96 if (ne->value > (ne->thresh << 1)) {
97 noise_level = kHigh;
98 } else {
99 if (ne->value > ne->thresh)
100 noise_level = kMedium;
101 else if (ne->value > (ne->thresh >> 1))
102 noise_level = kLow;
103 else
104 noise_level = kLowLow;
105 }
106 return noise_level;
107 }
108
vp9_update_noise_estimate(VP9_COMP * const cpi)109 void vp9_update_noise_estimate(VP9_COMP *const cpi) {
110 const VP9_COMMON *const cm = &cpi->common;
111 NOISE_ESTIMATE *const ne = &cpi->noise_estimate;
112 const int low_res = (cm->width <= 352 && cm->height <= 288);
113 // Estimate of noise level every frame_period frames.
114 int frame_period = 8;
115 int thresh_consec_zeromv = 6;
116 int frame_counter = cm->current_video_frame;
117 // Estimate is between current source and last source.
118 YV12_BUFFER_CONFIG *last_source = cpi->Last_Source;
119 #if CONFIG_VP9_TEMPORAL_DENOISING
120 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) {
121 last_source = &cpi->denoiser.last_source;
122 // Tune these thresholds for different resolutions when denoising is
123 // enabled.
124 if (cm->width > 640 && cm->width <= 1920) {
125 thresh_consec_zeromv = 2;
126 }
127 }
128 #endif
129 ne->enabled = enable_noise_estimation(cpi);
130 if (cpi->svc.number_spatial_layers > 1)
131 frame_counter = cpi->svc.current_superframe;
132 if (!ne->enabled || frame_counter % frame_period != 0 ||
133 last_source == NULL ||
134 (cpi->svc.number_spatial_layers == 1 &&
135 (ne->last_w != cm->width || ne->last_h != cm->height))) {
136 #if CONFIG_VP9_TEMPORAL_DENOISING
137 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi))
138 copy_frame(&cpi->denoiser.last_source, cpi->Source);
139 #endif
140 if (last_source != NULL) {
141 ne->last_w = cm->width;
142 ne->last_h = cm->height;
143 }
144 return;
145 } else if (frame_counter > 60 && cpi->svc.num_encoded_top_layer > 1 &&
146 cpi->rc.frames_since_key > cpi->svc.number_spatial_layers &&
147 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1 &&
148 cpi->rc.avg_frame_low_motion < (low_res ? 60 : 40)) {
149 // Force noise estimation to 0 and denoiser off if content has high motion.
150 ne->level = kLowLow;
151 ne->count = 0;
152 ne->num_frames_estimate = 10;
153 #if CONFIG_VP9_TEMPORAL_DENOISING
154 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi) &&
155 cpi->svc.current_superframe > 1) {
156 vp9_denoiser_set_noise_level(cpi, ne->level);
157 copy_frame(&cpi->denoiser.last_source, cpi->Source);
158 }
159 #endif
160 return;
161 } else {
162 unsigned int bin_size = 100;
163 unsigned int hist[MAX_VAR_HIST_BINS] = { 0 };
164 unsigned int hist_avg[MAX_VAR_HIST_BINS];
165 unsigned int max_bin = 0;
166 unsigned int max_bin_count = 0;
167 unsigned int bin_cnt;
168 int bsize = BLOCK_16X16;
169 // Loop over sub-sample of 16x16 blocks of frame, and for blocks that have
170 // been encoded as zero/small mv at least x consecutive frames, compute
171 // the variance to update estimate of noise in the source.
172 const uint8_t *src_y = cpi->Source->y_buffer;
173 const int src_ystride = cpi->Source->y_stride;
174 const uint8_t *last_src_y = last_source->y_buffer;
175 const int last_src_ystride = last_source->y_stride;
176 const uint8_t *src_u = cpi->Source->u_buffer;
177 const uint8_t *src_v = cpi->Source->v_buffer;
178 const int src_uvstride = cpi->Source->uv_stride;
179 int mi_row, mi_col;
180 int num_low_motion = 0;
181 int frame_low_motion = 1;
182 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
183 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
184 int bl_index = mi_row * cm->mi_cols + mi_col;
185 if (cpi->consec_zero_mv[bl_index] > thresh_consec_zeromv)
186 num_low_motion++;
187 }
188 }
189 if (num_low_motion < ((3 * cm->mi_rows * cm->mi_cols) >> 3))
190 frame_low_motion = 0;
191 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
192 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
193 // 16x16 blocks, 1/4 sample of frame.
194 if (mi_row % 4 == 0 && mi_col % 4 == 0 && mi_row < cm->mi_rows - 1 &&
195 mi_col < cm->mi_cols - 1) {
196 int bl_index = mi_row * cm->mi_cols + mi_col;
197 int bl_index1 = bl_index + 1;
198 int bl_index2 = bl_index + cm->mi_cols;
199 int bl_index3 = bl_index2 + 1;
200 int consec_zeromv =
201 VPXMIN(cpi->consec_zero_mv[bl_index],
202 VPXMIN(cpi->consec_zero_mv[bl_index1],
203 VPXMIN(cpi->consec_zero_mv[bl_index2],
204 cpi->consec_zero_mv[bl_index3])));
205 // Only consider blocks that are likely steady background. i.e, have
206 // been encoded as zero/low motion x (= thresh_consec_zeromv) frames
207 // in a row. consec_zero_mv[] defined for 8x8 blocks, so consider all
208 // 4 sub-blocks for 16x16 block. And exclude this frame if
209 // high_source_sad is true (i.e., scene/content change).
210 if (frame_low_motion && consec_zeromv > thresh_consec_zeromv &&
211 !cpi->rc.high_source_sad &&
212 !cpi->svc.high_source_sad_superframe) {
213 int is_skin = 0;
214 if (cpi->use_skin_detection) {
215 is_skin =
216 vp9_compute_skin_block(src_y, src_u, src_v, src_ystride,
217 src_uvstride, bsize, consec_zeromv, 0);
218 }
219 if (!is_skin) {
220 unsigned int sse;
221 // Compute variance between co-located blocks from current and
222 // last input frames.
223 unsigned int variance = cpi->fn_ptr[bsize].vf(
224 src_y, src_ystride, last_src_y, last_src_ystride, &sse);
225 unsigned int hist_index = variance / bin_size;
226 if (hist_index < MAX_VAR_HIST_BINS)
227 hist[hist_index]++;
228 else if (hist_index < 3 * (MAX_VAR_HIST_BINS >> 1))
229 hist[MAX_VAR_HIST_BINS - 1]++; // Account for the tail
230 }
231 }
232 }
233 src_y += 8;
234 last_src_y += 8;
235 src_u += 4;
236 src_v += 4;
237 }
238 src_y += (src_ystride << 3) - (cm->mi_cols << 3);
239 last_src_y += (last_src_ystride << 3) - (cm->mi_cols << 3);
240 src_u += (src_uvstride << 2) - (cm->mi_cols << 2);
241 src_v += (src_uvstride << 2) - (cm->mi_cols << 2);
242 }
243 ne->last_w = cm->width;
244 ne->last_h = cm->height;
245 // Adjust histogram to account for effect that histogram flattens
246 // and shifts to zero as scene darkens.
247 if (hist[0] > 10 && (hist[MAX_VAR_HIST_BINS - 1] > hist[0] >> 2)) {
248 hist[0] = 0;
249 hist[1] >>= 2;
250 hist[2] >>= 2;
251 hist[3] >>= 2;
252 hist[4] >>= 1;
253 hist[5] >>= 1;
254 hist[6] = 3 * hist[6] >> 1;
255 hist[MAX_VAR_HIST_BINS - 1] >>= 1;
256 }
257
258 // Average hist[] and find largest bin
259 for (bin_cnt = 0; bin_cnt < MAX_VAR_HIST_BINS; bin_cnt++) {
260 if (bin_cnt == 0)
261 hist_avg[bin_cnt] = (hist[0] + hist[1] + hist[2]) / 3;
262 else if (bin_cnt == MAX_VAR_HIST_BINS - 1)
263 hist_avg[bin_cnt] = hist[MAX_VAR_HIST_BINS - 1] >> 2;
264 else if (bin_cnt == MAX_VAR_HIST_BINS - 2)
265 hist_avg[bin_cnt] = (hist[bin_cnt - 1] + 2 * hist[bin_cnt] +
266 (hist[bin_cnt + 1] >> 1) + 2) >>
267 2;
268 else
269 hist_avg[bin_cnt] =
270 (hist[bin_cnt - 1] + 2 * hist[bin_cnt] + hist[bin_cnt + 1] + 2) >>
271 2;
272
273 if (hist_avg[bin_cnt] > max_bin_count) {
274 max_bin_count = hist_avg[bin_cnt];
275 max_bin = bin_cnt;
276 }
277 }
278
279 // Scale by 40 to work with existing thresholds
280 ne->value = (int)((3 * ne->value + max_bin * 40) >> 2);
281 // Quickly increase VNR strength when the noise level increases suddenly.
282 if (ne->level < kMedium && ne->value > ne->adapt_thresh) {
283 ne->count = ne->num_frames_estimate;
284 } else {
285 ne->count++;
286 }
287 if (ne->count == ne->num_frames_estimate) {
288 // Reset counter and check noise level condition.
289 ne->num_frames_estimate = 30;
290 ne->count = 0;
291 ne->level = vp9_noise_estimate_extract_level(ne);
292 #if CONFIG_VP9_TEMPORAL_DENOISING
293 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi))
294 vp9_denoiser_set_noise_level(cpi, ne->level);
295 #endif
296 }
297 }
298 #if CONFIG_VP9_TEMPORAL_DENOISING
299 if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi))
300 copy_frame(&cpi->denoiser.last_source, cpi->Source);
301 #endif
302 }
303