/* * 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.
*/
// TODO(jackychen): If increase_denoising is enabled in the future, // we might need to update the code for calculating 'total_adj' in // case the C code is not bit-exact with corresponding sse2 code. int vp9_denoiser_filter_c(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, int motion_magnitude) { int r, c; const uint8_t *sig_start = sig; const uint8_t *mc_avg_start = mc_avg;
uint8_t *avg_start = avg; int diff, adj, absdiff, delta; int adj_val[] = { 3, 4, 6 }; int total_adj = 0; int shift_inc = 1;
// If motion_magnitude is small, making the denoiser more aggressive by // increasing the adjustment for each level. Add another increment for // blocks that are labeled for increase denoising. if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) { if (increase_denoising) {
shift_inc = 2;
}
adj_val[0] += shift_inc;
adj_val[1] += shift_inc;
adj_val[2] += shift_inc;
}
// First attempt to apply a strong temporal denoising filter. for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) { for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
diff = mc_avg[c] - sig[c];
absdiff = abs(diff);
if (absdiff <= absdiff_thresh(bs, increase_denoising)) {
avg[c] = mc_avg[c];
total_adj += diff;
} else { switch (absdiff) { case 4: case 5: case 6: case 7: adj = adj_val[0]; break; case 8: case 9: case 10: case 11: case 12: case 13: case 14: case 15: adj = adj_val[1]; break; default: adj = adj_val[2];
} if (diff > 0) {
avg[c] = VPXMIN(UINT8_MAX, sig[c] + adj);
total_adj += adj;
} else {
avg[c] = VPXMAX(0, sig[c] - adj);
total_adj -= adj;
}
}
}
sig += sig_stride;
avg += avg_stride;
mc_avg += mc_avg_stride;
}
// 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)) >>
num_pels_log2_lookup[bs]) +
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 < (4 << b_height_log2_lookup[bs]); ++r) { for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
diff = mc_avg[c] - sig[c];
adj = abs(diff); if (adj > delta) {
adj = delta;
} if (diff > 0) { // Diff positive means we made positive adjustment above // (in first try/attempt), so now make negative adjustment to bring // denoised signal down.
avg[c] = VPXMAX(0, avg[c] - adj);
total_adj -= adj;
} else { // Diff negative means we made negative adjustment above // (in first try/attempt), so now make positive adjustment to bring // denoised signal up.
avg[c] = VPXMIN(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;
}
static uint8_t *block_start(uint8_t *framebuf, int stride, int mi_row, int mi_col) { return framebuf + (stride * mi_row << 3) + (mi_col << 3);
}
static VP9_DENOISER_DECISION perform_motion_compensation(
VP9_COMMON *const cm, VP9_DENOISER *denoiser, MACROBLOCK *mb, BLOCK_SIZE bs, int increase_denoising, int mi_row, int mi_col, PICK_MODE_CONTEXT *ctx, int motion_magnitude, int is_skin, int *zeromv_filter, int consec_zeromv, int num_spatial_layers, int width, int lst_fb_idx, int gld_fb_idx, int use_svc, int spatial_layer, int use_gf_temporal_ref) { constint sse_diff = (ctx->newmv_sse == UINT_MAX)
? 0
: ((int)ctx->zeromv_sse - (int)ctx->newmv_sse); int frame; int denoise_layer_idx = 0;
MACROBLOCKD *filter_mbd = &mb->e_mbd;
MODE_INFO *mi = filter_mbd->mi[0];
MODE_INFO saved_mi; int i; struct buf_2d saved_dst[MAX_MB_PLANE]; struct buf_2d saved_pre[MAX_MB_PLANE]; const RefBuffer *saved_block_refs[2];
MV_REFERENCE_FRAME saved_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 && frame != ALTREF_FRAME && frame != GOLDEN_FRAME &&
sse_diff > sse_diff_thresh(bs, increase_denoising, motion_magnitude)) {
mi->ref_frame[0] = ctx->best_reference_frame;
mi->mode = ctx->best_sse_inter_mode;
mi->mv[0] = ctx->best_sse_mv;
} else { // Otherwise, use the zero reference frame.
frame = ctx->best_zeromv_reference_frame;
ctx->newmv_sse = ctx->zeromv_sse; // Bias to last reference. if ((num_spatial_layers > 1 && !use_gf_temporal_ref) ||
frame == ALTREF_FRAME ||
(frame == GOLDEN_FRAME && use_gf_temporal_ref) ||
(frame != LAST_FRAME &&
((ctx->zeromv_lastref_sse < (5 * ctx->zeromv_sse) >> 2) ||
denoiser->denoising_level >= kDenHigh))) {
frame = LAST_FRAME;
ctx->newmv_sse = ctx->zeromv_lastref_sse;
}
mi->ref_frame[0] = frame;
mi->mode = ZEROMV;
mi->mv[0].as_int = 0;
ctx->best_sse_inter_mode = ZEROMV;
ctx->best_sse_mv.as_int = 0;
*zeromv_filter = 1; if (denoiser->denoising_level > kDenMedium) {
motion_magnitude = 0;
}
}
saved_frame = frame; // When using SVC, we need to map REF_FRAME to the frame buffer index. if (use_svc) { if (frame == LAST_FRAME)
frame = lst_fb_idx + 1; elseif (frame == GOLDEN_FRAME)
frame = gld_fb_idx + 1; // Shift for the second spatial layer. if (num_spatial_layers - spatial_layer == 2)
frame = frame + denoiser->num_ref_frames;
denoise_layer_idx = num_spatial_layers - spatial_layer - 1;
}
// Force copy (no denoise, copy source in denoised buffer) if // running_avg_y[frame] is NULL. if (denoiser->running_avg_y[frame].buffer_alloc == NULL) { // Restore everything to its original state
*mi = saved_mi; return COPY_BLOCK;
}
if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) { // Restore everything to its original state
*mi = saved_mi; return COPY_BLOCK;
} if (motion_magnitude > (noise_motion_thresh(bs, increase_denoising) << 3)) { // Restore everything to its original state
*mi = saved_mi; return COPY_BLOCK;
}
// We will restore these after motion compensation. for (i = 0; i < MAX_MB_PLANE; ++i) {
saved_pre[i] = filter_mbd->plane[i].pre[0];
saved_dst[i] = filter_mbd->plane[i].dst;
}
saved_block_refs[0] = filter_mbd->block_refs[0];
// 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;
// Restore everything to its original state
*mi = saved_mi;
filter_mbd->block_refs[0] = saved_block_refs[0]; for (i = 0; i < MAX_MB_PLANE; ++i) {
filter_mbd->plane[i].pre[0] = saved_pre[i];
filter_mbd->plane[i].dst = saved_dst[i];
}
return FILTER_BLOCK;
}
void vp9_denoiser_denoise(VP9_COMP *cpi, MACROBLOCK *mb, int mi_row, int mi_col,
BLOCK_SIZE bs, PICK_MODE_CONTEXT *ctx,
VP9_DENOISER_DECISION *denoiser_decision, int use_gf_temporal_ref) { int mv_col, mv_row; int motion_magnitude = 0; int zeromv_filter = 0;
VP9_DENOISER *denoiser = &cpi->denoiser;
VP9_DENOISER_DECISION decision = COPY_BLOCK;
if (cpi->use_skin_detection && bs <= BLOCK_32X32 &&
denoiser->denoising_level < kDenHigh) { int motion_level = (motion_magnitude < 16) ? 0 : 1; // If motion for current block is small/zero, compute consec_zeromv for // skin detection (early exit in skin detection is done for large // consec_zeromv when current block has small/zero motion).
consec_zeromv = 0; if (motion_level == 0) {
VP9_COMMON *const cm = &cpi->common; int j, i; // Loop through the 8x8 sub-blocks. constint bw = num_8x8_blocks_wide_lookup[bs]; constint bh = num_8x8_blocks_high_lookup[bs]; constint xmis = VPXMIN(cm->mi_cols - mi_col, bw); constint ymis = VPXMIN(cm->mi_rows - mi_row, bh); constint block_index = mi_row * cm->mi_cols + mi_col;
consec_zeromv = 100; for (i = 0; i < ymis; i++) { for (j = 0; j < xmis; j++) { int bl_index = block_index + i * cm->mi_cols + j;
consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index], consec_zeromv); // No need to keep checking 8x8 blocks if any of the sub-blocks // has small consec_zeromv (since threshold for no_skin based on // zero/small motion in skin detection is high, i.e., > 4). if (consec_zeromv < 4) {
i = ymis; break;
}
}
}
} // TODO(marpan): Compute skin detection over sub-blocks.
is_skin = vp9_compute_skin_block(
mb->plane[0].src.buf, mb->plane[1].src.buf, mb->plane[2].src.buf,
mb->plane[0].src.stride, mb->plane[1].src.stride, bs, consec_zeromv,
motion_level);
} if (!is_skin && denoiser->denoising_level == kDenHigh) increase_denoising = 1;
// Copy block if LAST_FRAME is not a reference. // Last doesn't always exist when SVC layers are dynamically changed, e.g. top // spatial layer doesn't have last reference when it's brought up for the // first time on the fly. if (last_is_reference && denoiser->denoising_level >= kDenLow &&
!ctx->sb_skip_denoising)
decision = perform_motion_compensation(
&cpi->common, denoiser, mb, bs, increase_denoising, mi_row, mi_col, ctx,
motion_magnitude, is_skin, &zeromv_filter, consec_zeromv,
cpi->svc.number_spatial_layers, cpi->Source->y_width, cpi->lst_fb_idx,
cpi->gld_fb_idx, cpi->use_svc, cpi->svc.spatial_layer_id,
use_gf_temporal_ref);
void vp9_denoiser_update_frame_info(
VP9_DENOISER *denoiser, YV12_BUFFER_CONFIG src, struct SVC *svc,
FRAME_TYPE frame_type, int refresh_alt_ref_frame, int refresh_golden_frame, int refresh_last_frame, int alt_fb_idx, int gld_fb_idx, int lst_fb_idx, int resized, int svc_refresh_denoiser_buffers, int second_spatial_layer) { constint shift = second_spatial_layer ? denoiser->num_ref_frames : 0; // Copy source into denoised reference buffers on KEY_FRAME or // if the just encoded frame was resized. For SVC, copy source if the base // spatial layer was key frame. if (frame_type == KEY_FRAME || resized != 0 || denoiser->reset ||
svc_refresh_denoiser_buffers) { int i; // Start at 1 so as not to overwrite the INTRA_FRAME for (i = 1; i < denoiser->num_ref_frames; ++i) { if (denoiser->running_avg_y[i + shift].buffer_alloc != NULL)
copy_frame(&denoiser->running_avg_y[i + shift], &src);
}
denoiser->reset = 0; return;
}
if (svc->temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS &&
svc->use_set_ref_frame_config) { int i; for (i = 0; i < REF_FRAMES; i++) { if (svc->update_buffer_slot[svc->spatial_layer_id] & (1 << i))
copy_frame(&denoiser->running_avg_y[i + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
}
} else { // If more than one refresh occurs, must copy frame buffer. if ((refresh_alt_ref_frame + refresh_golden_frame + refresh_last_frame) >
1) { if (refresh_alt_ref_frame) {
copy_frame(&denoiser->running_avg_y[alt_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
} if (refresh_golden_frame) {
copy_frame(&denoiser->running_avg_y[gld_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
} if (refresh_last_frame) {
copy_frame(&denoiser->running_avg_y[lst_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
}
} else { if (refresh_alt_ref_frame) {
swap_frame_buffer(&denoiser->running_avg_y[alt_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
} if (refresh_golden_frame) {
swap_frame_buffer(&denoiser->running_avg_y[gld_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
} if (refresh_last_frame) {
swap_frame_buffer(&denoiser->running_avg_y[lst_fb_idx + 1 + shift],
&denoiser->running_avg_y[INTRA_FRAME + shift]);
}
}
}
}
int vp9_denoiser_realloc_svc(VP9_COMMON *cm, VP9_DENOISER *denoiser, struct SVC *svc, int svc_buf_shift, int refresh_alt, int refresh_gld, int refresh_lst, int alt_fb_idx, int gld_fb_idx, int lst_fb_idx) { int fail = 0; if (svc->temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS &&
svc->use_set_ref_frame_config) { int i; for (i = 0; i < REF_FRAMES; i++) { if (cm->frame_type == KEY_FRAME ||
svc->update_buffer_slot[svc->spatial_layer_id] & (1 << i)) {
fail = vp9_denoiser_realloc_svc_helper(cm, denoiser,
i + 1 + svc_buf_shift);
}
}
} else { if (refresh_alt) { // Increase the frame buffer index by 1 to map it to the buffer index in // the denoiser.
fail = vp9_denoiser_realloc_svc_helper(cm, denoiser,
alt_fb_idx + 1 + svc_buf_shift); if (fail) return 1;
} if (refresh_gld) {
fail = vp9_denoiser_realloc_svc_helper(cm, denoiser,
gld_fb_idx + 1 + svc_buf_shift); if (fail) return 1;
} if (refresh_lst) {
fail = vp9_denoiser_realloc_svc_helper(cm, denoiser,
lst_fb_idx + 1 + svc_buf_shift); if (fail) return 1;
}
} return 0;
}
int vp9_denoiser_alloc(VP9_COMMON *cm, struct SVC *svc, VP9_DENOISER *denoiser, int use_svc, int noise_sen, int width, int height, int ssx, int ssy, #if CONFIG_VP9_HIGHBITDEPTH int use_highbitdepth, #endif int border) { int i, layer, fail, init_num_ref_frames; constint legacy_byte_alignment = 0; int num_layers = 1; int scaled_width = width; int scaled_height = height; if (use_svc) {
LAYER_CONTEXT *lc = &svc->layer_context[svc->spatial_layer_id *
svc->number_temporal_layers +
svc->temporal_layer_id];
get_layer_resolution(width, height, lc->scaling_factor_num,
lc->scaling_factor_den, &scaled_width, &scaled_height); // For SVC: only denoise at most 2 spatial (highest) layers. if (noise_sen >= 2) // Denoise from one spatial layer below the top.
svc->first_layer_denoise = VPXMAX(svc->number_spatial_layers - 2, 0); else // Only denoise the top spatial layer.
svc->first_layer_denoise = VPXMAX(svc->number_spatial_layers - 1, 0);
num_layers = svc->number_spatial_layers - svc->first_layer_denoise;
}
assert(denoiser != NULL);
denoiser->num_ref_frames = use_svc ? SVC_REF_FRAMES : NONSVC_REF_FRAMES;
init_num_ref_frames = use_svc ? MAX_REF_FRAMES : NONSVC_REF_FRAMES;
denoiser->num_layers = num_layers;
CHECK_MEM_ERROR(&cm->error, denoiser->running_avg_y,
vpx_calloc(denoiser->num_ref_frames * num_layers, sizeof(denoiser->running_avg_y[0])));
CHECK_MEM_ERROR(
&cm->error, denoiser->mc_running_avg_y,
vpx_calloc(num_layers, sizeof(denoiser->mc_running_avg_y[0])));
staticvoid force_refresh_longterm_ref(VP9_COMP *const cpi) {
SVC *const svc = &cpi->svc; // If long term reference is used, force refresh of that slot, so // denoiser buffer for long term reference stays in sync. if (svc->use_gf_temporal_ref_current_layer) { int index = svc->spatial_layer_id; if (svc->number_spatial_layers == 3) index = svc->spatial_layer_id - 1;
assert(index >= 0);
cpi->alt_fb_idx = svc->buffer_gf_temporal_ref[index].idx;
cpi->refresh_alt_ref_frame = 1;
}
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.