Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 1 | /* |
Krishna Rapaka | 7319db5 | 2021-09-28 20:35:29 -0700 | [diff] [blame] | 2 | * Copyright (c) 2021, Alliance for Open Media. All rights reserved |
Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 3 | * |
Vibhoothi | 41c6dd7 | 2021-10-12 18:48:26 +0000 | [diff] [blame] | 4 | * This source code is subject to the terms of the BSD 3-Clause Clear License |
| 5 | * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear |
| 6 | * License was not distributed with this source code in the LICENSE file, you |
| 7 | * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the |
| 8 | * Alliance for Open Media Patent License 1.0 was not distributed with this |
| 9 | * source code in the PATENTS file, you can obtain it at |
| 10 | * aomedia.org/license/patent-license/. |
Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 11 | */ |
| 12 | |
| 13 | #include <assert.h> |
| 14 | #include <limits.h> |
| 15 | #include <math.h> |
| 16 | |
| 17 | #include "config/aom_dsp_rtcd.h" |
| 18 | #include "aom_dsp/aom_dsp_common.h" |
| 19 | #include "aom_scale/yv12config.h" |
| 20 | #include "aom/aom_integer.h" |
| 21 | #include "av1/encoder/context_tree.h" |
| 22 | #include "av1/encoder/av1_noise_estimate.h" |
| 23 | #include "av1/encoder/encoder.h" |
| 24 | |
| 25 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 26 | // For SVC: only do noise estimation on top spatial layer. |
| 27 | static INLINE int noise_est_svc(const struct AV1_COMP *const cpi) { |
| 28 | return (!cpi->use_svc || |
| 29 | (cpi->use_svc && |
| 30 | cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)); |
| 31 | } |
| 32 | #endif |
| 33 | |
| 34 | void av1_noise_estimate_init(NOISE_ESTIMATE *const ne, int width, int height) { |
| 35 | ne->enabled = 0; |
| 36 | ne->level = (width * height < 1280 * 720) ? kLowLow : kLow; |
| 37 | ne->value = 0; |
| 38 | ne->count = 0; |
| 39 | ne->thresh = 90; |
| 40 | ne->last_w = 0; |
| 41 | ne->last_h = 0; |
| 42 | if (width * height >= 1920 * 1080) { |
| 43 | ne->thresh = 200; |
| 44 | } else if (width * height >= 1280 * 720) { |
| 45 | ne->thresh = 140; |
| 46 | } else if (width * height >= 640 * 360) { |
| 47 | ne->thresh = 115; |
| 48 | } |
| 49 | ne->num_frames_estimate = 15; |
| 50 | ne->adapt_thresh = (3 * ne->thresh) >> 1; |
| 51 | } |
| 52 | |
| 53 | static int enable_noise_estimation(AV1_COMP *const cpi) { |
| 54 | ResizePendingParams *const resize_pending_params = |
| 55 | &cpi->resize_pending_params; |
| 56 | const int resize_pending = |
| 57 | (resize_pending_params->width && resize_pending_params->height && |
| 58 | (cpi->common.width != resize_pending_params->width || |
| 59 | cpi->common.height != resize_pending_params->height)); |
| 60 | |
Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 61 | if (cpi->common.seq_params.use_highbitdepth) return 0; |
Ryan Lei | 2a67637 | 2020-09-28 16:58:52 -0700 | [diff] [blame] | 62 | |
Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 63 | // Enable noise estimation if denoising is on. |
| 64 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 65 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi) && |
| 66 | cpi->common.width >= 320 && cpi->common.height >= 180) |
| 67 | return 1; |
| 68 | #endif |
| 69 | // Only allow noise estimate under certain encoding mode. |
| 70 | // Enabled for 1 pass CBR, speed >=5, and if resolution is same as original. |
| 71 | // Not enabled for SVC mode and screen_content_mode. |
| 72 | // Not enabled for low resolutions. |
| 73 | if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_cfg.mode == AOM_CBR && |
| 74 | cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && cpi->oxcf.speed >= 5 && |
| 75 | resize_pending == 0 && !cpi->use_svc && |
Vishesh | 94a6529 | 2020-07-01 15:28:53 +0530 | [diff] [blame] | 76 | cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN && |
Fyodor Kyslov | 34b591f | 2020-06-24 20:13:58 -0700 | [diff] [blame] | 77 | cpi->common.width * cpi->common.height >= 640 * 360) |
| 78 | return 1; |
| 79 | else |
| 80 | return 0; |
| 81 | } |
| 82 | |
| 83 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 84 | static void copy_frame(YV12_BUFFER_CONFIG *const dest, |
| 85 | const YV12_BUFFER_CONFIG *const src) { |
| 86 | const uint8_t *srcbuf = src->y_buffer; |
| 87 | uint8_t *destbuf = dest->y_buffer; |
| 88 | |
| 89 | assert(dest->y_width == src->y_width); |
| 90 | assert(dest->y_height == src->y_height); |
| 91 | |
| 92 | for (int r = 0; r < dest->y_height; ++r) { |
| 93 | memcpy(destbuf, srcbuf, dest->y_width); |
| 94 | destbuf += dest->y_stride; |
| 95 | srcbuf += src->y_stride; |
| 96 | } |
| 97 | } |
| 98 | #endif // CONFIG_AV1_TEMPORAL_DENOISING |
| 99 | |
| 100 | NOISE_LEVEL av1_noise_estimate_extract_level(NOISE_ESTIMATE *const ne) { |
| 101 | int noise_level = kLowLow; |
| 102 | if (ne->value > (ne->thresh << 1)) { |
| 103 | noise_level = kHigh; |
| 104 | } else { |
| 105 | if (ne->value > ne->thresh) |
| 106 | noise_level = kMedium; |
| 107 | else if (ne->value > (ne->thresh >> 1)) |
| 108 | noise_level = kLow; |
| 109 | else |
| 110 | noise_level = kLowLow; |
| 111 | } |
| 112 | return noise_level; |
| 113 | } |
| 114 | |
| 115 | void av1_update_noise_estimate(AV1_COMP *const cpi) { |
| 116 | const AV1_COMMON *const cm = &cpi->common; |
| 117 | const CommonModeInfoParams *const mi_params = &cm->mi_params; |
| 118 | |
| 119 | NOISE_ESTIMATE *const ne = &cpi->noise_estimate; |
| 120 | const int low_res = (cm->width <= 352 && cm->height <= 288); |
| 121 | // Estimate of noise level every frame_period frames. |
| 122 | int frame_period = 8; |
| 123 | int thresh_consec_zeromv = 6; |
| 124 | int frame_counter = cm->current_frame.frame_number; |
| 125 | // Estimate is between current source and last source. |
| 126 | YV12_BUFFER_CONFIG *last_source = cpi->last_source; |
| 127 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 128 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) { |
| 129 | last_source = &cpi->denoiser.last_source; |
| 130 | // Tune these thresholds for different resolutions when denoising is |
| 131 | // enabled. |
| 132 | if (cm->width > 640 && cm->width <= 1920) { |
| 133 | thresh_consec_zeromv = 2; |
| 134 | } |
| 135 | } |
| 136 | #endif |
| 137 | ne->enabled = enable_noise_estimation(cpi); |
| 138 | if (cpi->svc.number_spatial_layers > 1) |
| 139 | frame_counter = cpi->svc.current_superframe; |
| 140 | if (!ne->enabled || frame_counter % frame_period != 0 || |
| 141 | last_source == NULL || |
| 142 | (cpi->svc.number_spatial_layers == 1 && |
| 143 | (ne->last_w != cm->width || ne->last_h != cm->height))) { |
| 144 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 145 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) |
| 146 | copy_frame(&cpi->denoiser.last_source, cpi->source); |
| 147 | #endif |
| 148 | if (last_source != NULL) { |
| 149 | ne->last_w = cm->width; |
| 150 | ne->last_h = cm->height; |
| 151 | } |
| 152 | return; |
| 153 | } else if (frame_counter > 60 && cpi->svc.num_encoded_top_layer > 1 && |
| 154 | cpi->rc.frames_since_key > cpi->svc.number_spatial_layers && |
| 155 | cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1 && |
| 156 | cpi->rc.avg_frame_low_motion < (low_res ? 60 : 40)) { |
| 157 | // Force noise estimation to 0 and denoiser off if content has high motion. |
| 158 | ne->level = kLowLow; |
| 159 | ne->count = 0; |
| 160 | ne->num_frames_estimate = 10; |
| 161 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 162 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi) && |
| 163 | cpi->svc.current_superframe > 1) { |
| 164 | av1_denoiser_set_noise_level(cpi, ne->level); |
| 165 | copy_frame(&cpi->denoiser.last_source, cpi->source); |
| 166 | } |
| 167 | #endif |
| 168 | return; |
| 169 | } else { |
| 170 | unsigned int bin_size = 100; |
| 171 | unsigned int hist[MAX_VAR_HIST_BINS] = { 0 }; |
| 172 | unsigned int hist_avg[MAX_VAR_HIST_BINS]; |
| 173 | unsigned int max_bin = 0; |
| 174 | unsigned int max_bin_count = 0; |
| 175 | unsigned int bin_cnt; |
| 176 | int bsize = BLOCK_16X16; |
| 177 | // Loop over sub-sample of 16x16 blocks of frame, and for blocks that have |
| 178 | // been encoded as zero/small mv at least x consecutive frames, compute |
| 179 | // the variance to update estimate of noise in the source. |
| 180 | const uint8_t *src_y = cpi->source->y_buffer; |
| 181 | const int src_ystride = cpi->source->y_stride; |
| 182 | const uint8_t *last_src_y = last_source->y_buffer; |
| 183 | const int last_src_ystride = last_source->y_stride; |
| 184 | const uint8_t *src_u = cpi->source->u_buffer; |
| 185 | const uint8_t *src_v = cpi->source->v_buffer; |
| 186 | const int src_uvstride = cpi->source->uv_stride; |
| 187 | int mi_row, mi_col; |
| 188 | int num_low_motion = 0; |
| 189 | int frame_low_motion = 1; |
| 190 | for (mi_row = 0; mi_row < mi_params->mi_rows; mi_row += 2) { |
| 191 | for (mi_col = 0; mi_col < mi_params->mi_cols; mi_col += 2) { |
| 192 | int bl_index = |
| 193 | (mi_row >> 1) * (mi_params->mi_cols >> 1) + (mi_col >> 1); |
| 194 | if (cpi->consec_zero_mv[bl_index] > thresh_consec_zeromv) |
| 195 | num_low_motion++; |
| 196 | } |
| 197 | } |
| 198 | if (num_low_motion < |
| 199 | (((3 * (mi_params->mi_rows * mi_params->mi_cols) >> 2)) >> 3)) |
| 200 | frame_low_motion = 0; |
| 201 | for (mi_row = 0; mi_row < mi_params->mi_rows; mi_row++) { |
| 202 | for (mi_col = 0; mi_col < mi_params->mi_cols; mi_col++) { |
| 203 | // 16x16 blocks, 1/4 sample of frame. |
| 204 | if (mi_row % 8 == 0 && mi_col % 8 == 0 && |
| 205 | mi_row < mi_params->mi_rows - 3 && |
| 206 | mi_col < mi_params->mi_cols - 3) { |
| 207 | int bl_index = |
| 208 | (mi_row >> 1) * (mi_params->mi_cols >> 1) + (mi_col >> 1); |
| 209 | int bl_index1 = bl_index + 1; |
| 210 | int bl_index2 = bl_index + (mi_params->mi_cols >> 1); |
| 211 | int bl_index3 = bl_index2 + 1; |
| 212 | int consec_zeromv = |
| 213 | AOMMIN(cpi->consec_zero_mv[bl_index], |
| 214 | AOMMIN(cpi->consec_zero_mv[bl_index1], |
| 215 | AOMMIN(cpi->consec_zero_mv[bl_index2], |
| 216 | cpi->consec_zero_mv[bl_index3]))); |
| 217 | // Only consider blocks that are likely steady background. i.e, have |
| 218 | // been encoded as zero/low motion x (= thresh_consec_zeromv) frames |
| 219 | // in a row. consec_zero_mv[] defined for 8x8 blocks, so consider all |
| 220 | // 4 sub-blocks for 16x16 block. And exclude this frame if |
| 221 | // high_source_sad is true (i.e., scene/content change). |
| 222 | if (frame_low_motion && consec_zeromv > thresh_consec_zeromv && |
| 223 | !cpi->rc.high_source_sad) { |
| 224 | unsigned int sse; |
| 225 | // Compute variance between co-located blocks from current and |
| 226 | // last input frames. |
| 227 | unsigned int variance = cpi->fn_ptr[bsize].vf( |
| 228 | src_y, src_ystride, last_src_y, last_src_ystride, &sse); |
| 229 | unsigned int hist_index = variance / bin_size; |
| 230 | if (hist_index < MAX_VAR_HIST_BINS) |
| 231 | hist[hist_index]++; |
| 232 | else if (hist_index < 3 * (MAX_VAR_HIST_BINS >> 1)) |
| 233 | hist[MAX_VAR_HIST_BINS - 1]++; // Account for the tail |
| 234 | } |
| 235 | } |
| 236 | src_y += 4; |
| 237 | last_src_y += 4; |
| 238 | src_u += 2; |
| 239 | src_v += 2; |
| 240 | } |
| 241 | src_y += (src_ystride << 2) - (mi_params->mi_cols << 2); |
| 242 | last_src_y += (last_src_ystride << 2) - (mi_params->mi_cols << 2); |
| 243 | src_u += (src_uvstride << 1) - (mi_params->mi_cols << 1); |
| 244 | src_v += (src_uvstride << 1) - (mi_params->mi_cols << 1); |
| 245 | } |
| 246 | ne->last_w = cm->width; |
| 247 | ne->last_h = cm->height; |
| 248 | // Adjust histogram to account for effect that histogram flattens |
| 249 | // and shifts to zero as scene darkens. |
| 250 | if (hist[0] > 10 && (hist[MAX_VAR_HIST_BINS - 1] > hist[0] >> 2)) { |
| 251 | hist[0] = 0; |
| 252 | hist[1] >>= 2; |
| 253 | hist[2] >>= 2; |
| 254 | hist[3] >>= 2; |
| 255 | hist[4] >>= 1; |
| 256 | hist[5] >>= 1; |
| 257 | hist[6] = 3 * hist[6] >> 1; |
| 258 | hist[MAX_VAR_HIST_BINS - 1] >>= 1; |
| 259 | } |
| 260 | |
| 261 | // Average hist[] and find largest bin |
| 262 | for (bin_cnt = 0; bin_cnt < MAX_VAR_HIST_BINS; bin_cnt++) { |
| 263 | if (bin_cnt == 0) |
| 264 | hist_avg[bin_cnt] = (hist[0] + hist[1] + hist[2]) / 3; |
| 265 | else if (bin_cnt == MAX_VAR_HIST_BINS - 1) |
| 266 | hist_avg[bin_cnt] = hist[MAX_VAR_HIST_BINS - 1] >> 2; |
| 267 | else if (bin_cnt == MAX_VAR_HIST_BINS - 2) |
| 268 | hist_avg[bin_cnt] = (hist[bin_cnt - 1] + 2 * hist[bin_cnt] + |
| 269 | (hist[bin_cnt + 1] >> 1) + 2) >> |
| 270 | 2; |
| 271 | else |
| 272 | hist_avg[bin_cnt] = |
| 273 | (hist[bin_cnt - 1] + 2 * hist[bin_cnt] + hist[bin_cnt + 1] + 2) >> |
| 274 | 2; |
| 275 | |
| 276 | if (hist_avg[bin_cnt] > max_bin_count) { |
| 277 | max_bin_count = hist_avg[bin_cnt]; |
| 278 | max_bin = bin_cnt; |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | // Scale by 40 to work with existing thresholds |
| 283 | ne->value = (int)((3 * ne->value + max_bin * 40) >> 2); |
| 284 | // Quickly increase VNR strength when the noise level increases suddenly. |
| 285 | if (ne->level < kMedium && ne->value > ne->adapt_thresh) { |
| 286 | ne->count = ne->num_frames_estimate; |
| 287 | } else { |
| 288 | ne->count++; |
| 289 | } |
| 290 | if (ne->count == ne->num_frames_estimate) { |
| 291 | // Reset counter and check noise level condition. |
| 292 | ne->num_frames_estimate = 30; |
| 293 | ne->count = 0; |
| 294 | ne->level = av1_noise_estimate_extract_level(ne); |
| 295 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 296 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) |
| 297 | av1_denoiser_set_noise_level(cpi, ne->level); |
| 298 | #endif |
| 299 | } |
| 300 | } |
| 301 | #if CONFIG_AV1_TEMPORAL_DENOISING |
| 302 | if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) |
| 303 | copy_frame(&cpi->denoiser.last_source, cpi->source); |
| 304 | #endif |
| 305 | } |