/* * 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. */ #include <limits.h> #include "denoising.h" #include "vp8/common/reconinter.h"#include "vpx/vpx_integer.h"#include "vpx_mem/vpx_mem.h"#include "vp8_rtcd.h" static const unsigned int NOISE_MOTION_THRESHOLD = 25 * 25;/* SSE_DIFF_THRESHOLD is selected as ~95% confidence assuming * var(noise) ~= 100. */static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20;static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 80; /* * The filter function was modified to reduce the computational complexity. * Step 1: * Instead of applying tap coefficients for each pixel, we calculated the * pixel adjustments vs. pixel diff value ahead of time. * adjustment = filtered_value - current_raw * = (filter_coefficient * diff + 128) >> 8 * where * filter_coefficient = (255 << 8) / (256 + ((absdiff * 330) >> 3)); * filter_coefficient += filter_coefficient / * (3 + motion_magnitude_adjustment); * filter_coefficient is clamped to 0 ~ 255. * * Step 2: * The adjustment vs. diff curve becomes flat very quick when diff increases. * This allowed us to use only several levels to approximate the curve without * changing the filtering algorithm too much. * The adjustments were further corrected by checking the motion magnitude. * The levels used are: * diff adjustment w/o motion correction adjustment w/ motion correction * [-255, -16] -6 -7 * [-15, -8] -4 -5 * [-7, -4] -3 -4 * [-3, 3] diff diff * [4, 7] 3 4 * [8, 15] 4 5 * [16, 255] 6 7 */ int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising) { unsigned char *running_avg_y_start = running_avg_y; unsigned char *sig_start = sig; int sum_diff_thresh; int r, c; int sum_diff = 0; int adj_val[3] = { 3, 4, 6 }; int shift_inc1 = 0; int shift_inc2 = 1; int col_sum[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }; /* 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_inc1 = 1; shift_inc2 = 2; } adj_val[0] += shift_inc2; adj_val[1] += shift_inc2; adj_val[2] += shift_inc2; } for (r = 0; r < 16; ++r) { for (c = 0; c < 16; ++c) { int diff = 0; int adjustment = 0; int absdiff = 0; diff = mc_running_avg_y[c] - sig[c]; absdiff = abs(diff); // When |diff| <= |3 + shift_inc1|, use pixel value from // last denoised raw. if (absdiff <= 3 + shift_inc1) { running_avg_y[c] = mc_running_avg_y[c]; col_sum[c] += diff; } else { if (absdiff >= 4 + shift_inc1 && absdiff <= 7) { adjustment = adj_val[0]; } else if (absdiff >= 8 && absdiff <= 15) { adjustment = adj_val[1]; } else { adjustment = adj_val[2]; } if (diff > 0) { if ((sig[c] + adjustment) > 255) { running_avg_y[c] = 255; } else { running_avg_y[c] = sig[c] + adjustment; } col_sum[c] += adjustment; } else { if ((sig[c] - adjustment) < 0) { running_avg_y[c] = 0; } else { running_avg_y[c] = sig[c] - adjustment; } col_sum[c] -= adjustment; } } } /* Update pointers for next iteration. */ sig += sig_stride; mc_running_avg_y += mc_avg_y_stride; running_avg_y += avg_y_stride; } for (c = 0; c < 16; ++c) { // Below we clip the value in the same way which SSE code use. // When adopting aggressive denoiser, the adj_val for each pixel // could be at most 8 (this is current max adjustment of the map). // In SSE code, we calculate the sum of adj_val for // the columns, so the sum could be upto 128(16 rows). However, // the range of the value is -128 ~ 127 in SSE code, that's why // we do this change in C code. // We don't do this for UV denoiser, since there are only 8 rows, // and max adjustments <= 8, so the sum of the columns will not // exceed 64. if (col_sum[c] >= 128) { col_sum[c] = 127; } sum_diff += col_sum[c]; } sum_diff_thresh = SUM_DIFF_THRESHOLD; if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH; if (abs(sum_diff) > sum_diff_thresh) { // Before returning to copy the block (i.e., apply no denoising), check // if we can still apply some (weaker) temporal filtering to this block, // that would otherwise not be denoised at all. Simplest is to apply // an additional adjustment to running_avg_y to bring it closer to sig. // The adjustment is capped by a maximum delta, and chosen such that // in most cases the resulting sum_diff will be within the // accceptable range given by sum_diff_thresh. // The delta is set by the excess of absolute pixel diff over threshold. int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; // Only apply the adjustment for max delta up to 3. if (delta < 4) { sig -= sig_stride * 16; mc_running_avg_y -= mc_avg_y_stride * 16; running_avg_y -= avg_y_stride * 16; for (r = 0; r < 16; ++r) { for (c = 0; c < 16; ++c) { int diff = mc_running_avg_y[c] - sig[c]; int adjustment = abs(diff); if (adjustment > delta) adjustment = delta; if (diff > 0) { // Bring denoised signal down. if (running_avg_y[c] - adjustment < 0) { running_avg_y[c] = 0; } else { running_avg_y[c] = running_avg_y[c] - adjustment; } col_sum[c] -= adjustment; } else if (diff < 0) { // Bring denoised signal up. if (running_avg_y[c] + adjustment > 255) { running_avg_y[c] = 255; } else { running_avg_y[c] = running_avg_y[c] + adjustment; } col_sum[c] += adjustment; } } // TODO(marpan): Check here if abs(sum_diff) has gone below the // threshold sum_diff_thresh, and if so, we can exit the row loop. sig += sig_stride; mc_running_avg_y += mc_avg_y_stride; running_avg_y += avg_y_stride; } sum_diff = 0; for (c = 0; c < 16; ++c) { if (col_sum[c] >= 128) { col_sum[c] = 127; } sum_diff += col_sum[c]; } if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; } else { return COPY_BLOCK; } } vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride); return FILTER_BLOCK;} int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising) { unsigned char *running_avg_start = running_avg; unsigned char *sig_start = sig; int sum_diff_thresh; int r, c; int sum_diff = 0; int sum_block = 0; int adj_val[3] = { 3, 4, 6 }; int shift_inc1 = 0; int shift_inc2 = 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_UV) { if (increase_denoising) { shift_inc1 = 1; shift_inc2 = 2; } adj_val[0] += shift_inc2; adj_val[1] += shift_inc2; adj_val[2] += shift_inc2; } // Avoid denoising color signal if its close to average level. for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { sum_block += sig[c]; } sig += sig_stride; } if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) { return COPY_BLOCK; } sig -= sig_stride * 8; for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { int diff = 0; int adjustment = 0; int absdiff = 0; diff = mc_running_avg[c] - sig[c]; absdiff = abs(diff); // When |diff| <= |3 + shift_inc1|, use pixel value from // last denoised raw. if (absdiff <= 3 + shift_inc1) { running_avg[c] = mc_running_avg[c]; sum_diff += diff; } else { if (absdiff >= 4 && absdiff <= 7) { adjustment = adj_val[0]; } else if (absdiff >= 8 && absdiff <= 15) { adjustment = adj_val[1]; } else { adjustment = adj_val[2]; } if (diff > 0) { if ((sig[c] + adjustment) > 255) { running_avg[c] = 255; } else { running_avg[c] = sig[c] + adjustment; } sum_diff += adjustment; } else { if ((sig[c] - adjustment) < 0) { running_avg[c] = 0; } else { running_avg[c] = sig[c] - adjustment; } sum_diff -= adjustment; } } } /* Update pointers for next iteration. */ sig += sig_stride; mc_running_avg += mc_avg_stride; running_avg += avg_stride; } sum_diff_thresh = SUM_DIFF_THRESHOLD_UV; if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV; if (abs(sum_diff) > sum_diff_thresh) { // Before returning to copy the block (i.e., apply no denoising), check // if we can still apply some (weaker) temporal filtering to this block, // that would otherwise not be denoised at all. Simplest is to apply // an additional adjustment to running_avg_y to bring it closer to sig. // The adjustment is capped by a maximum delta, and chosen such that // in most cases the resulting sum_diff will be within the // accceptable range given by sum_diff_thresh. // The delta is set by the excess of absolute pixel diff over threshold. int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; // Only apply the adjustment for max delta up to 3. if (delta < 4) { sig -= sig_stride * 8; mc_running_avg -= mc_avg_stride * 8; running_avg -= avg_stride * 8; for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { int diff = mc_running_avg[c] - sig[c]; int adjustment = abs(diff); if (adjustment > delta) adjustment = delta; if (diff > 0) { // Bring denoised signal down. if (running_avg[c] - adjustment < 0) { running_avg[c] = 0; } else { running_avg[c] = running_avg[c] - adjustment; } sum_diff -= adjustment; } else if (diff < 0) { // Bring denoised signal up. if (running_avg[c] + adjustment > 255) { running_avg[c] = 255; } else { running_avg[c] = running_avg[c] + adjustment; } sum_diff += adjustment; } } // TODO(marpan): Check here if abs(sum_diff) has gone below the // threshold sum_diff_thresh, and if so, we can exit the row loop. sig += sig_stride; mc_running_avg += mc_avg_stride; running_avg += avg_stride; } if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; } else { return COPY_BLOCK; } } vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride); return FILTER_BLOCK;} void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) { assert(mode > 0); // Denoiser is allocated only if mode > 0. if (mode == 1) { denoiser->denoiser_mode = kDenoiserOnYOnly; } else if (mode == 2) { denoiser->denoiser_mode = kDenoiserOnYUV; } else if (mode == 3) { denoiser->denoiser_mode = kDenoiserOnYUVAggressive; } else { denoiser->denoiser_mode = kDenoiserOnYUV; } if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) { denoiser->denoise_pars.scale_sse_thresh = 1; denoiser->denoise_pars.scale_motion_thresh = 8; denoiser->denoise_pars.scale_increase_filter = 0; denoiser->denoise_pars.denoise_mv_bias = 95; denoiser->denoise_pars.pickmode_mv_bias = 100; denoiser->denoise_pars.qp_thresh = 0; denoiser->denoise_pars.consec_zerolast = UINT_MAX; denoiser->denoise_pars.spatial_blur = 0; } else { denoiser->denoise_pars.scale_sse_thresh = 2; denoiser->denoise_pars.scale_motion_thresh = 16; denoiser->denoise_pars.scale_increase_filter = 1; denoiser->denoise_pars.denoise_mv_bias = 60; denoiser->denoise_pars.pickmode_mv_bias = 75; denoiser->denoise_pars.qp_thresh = 80; denoiser->denoise_pars.consec_zerolast = 15; denoiser->denoise_pars.spatial_blur = 0; }} int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height, int num_mb_rows, int num_mb_cols, int mode) { int i; assert(denoiser); denoiser->num_mb_cols = num_mb_cols; for (i = 0; i < MAX_REF_FRAMES; ++i) { denoiser->yv12_running_avg[i].flags = 0; if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_running_avg[i].buffer_alloc, 0, denoiser->yv12_running_avg[i].frame_size); } denoiser->yv12_mc_running_avg.flags = 0; if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0, denoiser->yv12_mc_running_avg.frame_size); if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_last_source.buffer_alloc, 0, denoiser->yv12_last_source.frame_size); denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1); if (!denoiser->denoise_state) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols)); vp8_denoiser_set_parameters(denoiser, mode); denoiser->nmse_source_diff = 0; denoiser->nmse_source_diff_count = 0; denoiser->qp_avg = 0; // QP threshold below which we can go up to aggressive mode. denoiser->qp_threshold_up = 80; // QP threshold above which we can go back down to normal mode. // For now keep this second threshold high, so not used currently. denoiser->qp_threshold_down = 128; // Bitrate thresholds and noise metric (nmse) thresholds for switching to // aggressive mode. // TODO(marpan): Adjust thresholds, including effect on resolution. denoiser->bitrate_threshold = 400000; // (bits/sec). denoiser->threshold_aggressive_mode = 80; if (width * height > 1280 * 720) { denoiser->bitrate_threshold = 3000000; denoiser->threshold_aggressive_mode = 200; } else if (width * height > 960 * 540) { denoiser->bitrate_threshold = 1200000; denoiser->threshold_aggressive_mode = 120; } else if (width * height > 640 * 480) { denoiser->bitrate_threshold = 600000; denoiser->threshold_aggressive_mode = 100; } return 0;} void vp8_denoiser_free(VP8_DENOISER *denoiser) { int i; assert(denoiser); for (i = 0; i < MAX_REF_FRAMES; ++i) { vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]); } vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg); vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source); vpx_free(denoiser->denoise_state);} void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser, MACROBLOCK *x, unsigned int best_sse, unsigned int zero_mv_sse, int recon_yoffset, int recon_uvoffset, loop_filter_info_n *lfi_n, int mb_row, int mb_col, int block_index, int consec_zero_last) { int mv_row; int mv_col; unsigned int motion_threshold; unsigned int motion_magnitude2; unsigned int sse_thresh; int sse_diff_thresh = 0; // Spatial loop filter: only applied selectively based on // temporal filter state of block relative to top/left neighbors. int apply_spatial_loop_filter = 1; MV_REFERENCE_FRAME frame = x->best_reference_frame; MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame; enum vp8_denoiser_decision decision = FILTER_BLOCK; enum vp8_denoiser_decision decision_u = COPY_BLOCK; enum vp8_denoiser_decision decision_v = COPY_BLOCK; if (zero_frame) { YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame]; YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg; YV12_BUFFER_CONFIG saved_pre, saved_dst; MB_MODE_INFO saved_mbmi; MACROBLOCKD *filter_xd = &x->e_mbd; MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi; int sse_diff = 0;Lines 1–500 of 725. Load more
/* * 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. */ #include <limits.h> #include "denoising.h" #include "vp8/common/reconinter.h" #include "vpx/vpx_integer.h" #include "vpx_mem/vpx_mem.h" #include "vp8_rtcd.h" static const unsigned int NOISE_MOTION_THRESHOLD = 25 * 25; /* SSE_DIFF_THRESHOLD is selected as ~95% confidence assuming * var(noise) ~= 100. */ static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20; static const unsigned int SSE_THRESHOLD = 16 * 16 * 40; static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 80; /* * The filter function was modified to reduce the computational complexity. * Step 1: * Instead of applying tap coefficients for each pixel, we calculated the * pixel adjustments vs. pixel diff value ahead of time. * adjustment = filtered_value - current_raw * = (filter_coefficient * diff + 128) >> 8 * where * filter_coefficient = (255 << 8) / (256 + ((absdiff * 330) >> 3)); * filter_coefficient += filter_coefficient / * (3 + motion_magnitude_adjustment); * filter_coefficient is clamped to 0 ~ 255. * * Step 2: * The adjustment vs. diff curve becomes flat very quick when diff increases. * This allowed us to use only several levels to approximate the curve without * changing the filtering algorithm too much. * The adjustments were further corrected by checking the motion magnitude. * The levels used are: * diff adjustment w/o motion correction adjustment w/ motion correction * [-255, -16] -6 -7 * [-15, -8] -4 -5 * [-7, -4] -3 -4 * [-3, 3] diff diff * [4, 7] 3 4 * [8, 15] 4 5 * [16, 255] 6 7 */ int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising) { unsigned char *running_avg_y_start = running_avg_y; unsigned char *sig_start = sig; int sum_diff_thresh; int r, c; int sum_diff = 0; int adj_val[3] = { 3, 4, 6 }; int shift_inc1 = 0; int shift_inc2 = 1; int col_sum[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }; /* 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_inc1 = 1; shift_inc2 = 2; } adj_val[0] += shift_inc2; adj_val[1] += shift_inc2; adj_val[2] += shift_inc2; } for (r = 0; r < 16; ++r) { for (c = 0; c < 16; ++c) { int diff = 0; int adjustment = 0; int absdiff = 0; diff = mc_running_avg_y[c] - sig[c]; absdiff = abs(diff); // When |diff| <= |3 + shift_inc1|, use pixel value from // last denoised raw. if (absdiff <= 3 + shift_inc1) { running_avg_y[c] = mc_running_avg_y[c]; col_sum[c] += diff; } else { if (absdiff >= 4 + shift_inc1 && absdiff <= 7) { adjustment = adj_val[0]; } else if (absdiff >= 8 && absdiff <= 15) { adjustment = adj_val[1]; } else { adjustment = adj_val[2]; } if (diff > 0) { if ((sig[c] + adjustment) > 255) { running_avg_y[c] = 255; } else { running_avg_y[c] = sig[c] + adjustment; } col_sum[c] += adjustment; } else { if ((sig[c] - adjustment) < 0) { running_avg_y[c] = 0; } else { running_avg_y[c] = sig[c] - adjustment; } col_sum[c] -= adjustment; } } } /* Update pointers for next iteration. */ sig += sig_stride; mc_running_avg_y += mc_avg_y_stride; running_avg_y += avg_y_stride; } for (c = 0; c < 16; ++c) { // Below we clip the value in the same way which SSE code use. // When adopting aggressive denoiser, the adj_val for each pixel // could be at most 8 (this is current max adjustment of the map). // In SSE code, we calculate the sum of adj_val for // the columns, so the sum could be upto 128(16 rows). However, // the range of the value is -128 ~ 127 in SSE code, that's why // we do this change in C code. // We don't do this for UV denoiser, since there are only 8 rows, // and max adjustments <= 8, so the sum of the columns will not // exceed 64. if (col_sum[c] >= 128) { col_sum[c] = 127; } sum_diff += col_sum[c]; } sum_diff_thresh = SUM_DIFF_THRESHOLD; if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH; if (abs(sum_diff) > sum_diff_thresh) { // Before returning to copy the block (i.e., apply no denoising), check // if we can still apply some (weaker) temporal filtering to this block, // that would otherwise not be denoised at all. Simplest is to apply // an additional adjustment to running_avg_y to bring it closer to sig. // The adjustment is capped by a maximum delta, and chosen such that // in most cases the resulting sum_diff will be within the // accceptable range given by sum_diff_thresh. // The delta is set by the excess of absolute pixel diff over threshold. int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; // Only apply the adjustment for max delta up to 3. if (delta < 4) { sig -= sig_stride * 16; mc_running_avg_y -= mc_avg_y_stride * 16; running_avg_y -= avg_y_stride * 16; for (r = 0; r < 16; ++r) { for (c = 0; c < 16; ++c) { int diff = mc_running_avg_y[c] - sig[c]; int adjustment = abs(diff); if (adjustment > delta) adjustment = delta; if (diff > 0) { // Bring denoised signal down. if (running_avg_y[c] - adjustment < 0) { running_avg_y[c] = 0; } else { running_avg_y[c] = running_avg_y[c] - adjustment; } col_sum[c] -= adjustment; } else if (diff < 0) { // Bring denoised signal up. if (running_avg_y[c] + adjustment > 255) { running_avg_y[c] = 255; } else { running_avg_y[c] = running_avg_y[c] + adjustment; } col_sum[c] += adjustment; } } // TODO(marpan): Check here if abs(sum_diff) has gone below the // threshold sum_diff_thresh, and if so, we can exit the row loop. sig += sig_stride; mc_running_avg_y += mc_avg_y_stride; running_avg_y += avg_y_stride; } sum_diff = 0; for (c = 0; c < 16; ++c) { if (col_sum[c] >= 128) { col_sum[c] = 127; } sum_diff += col_sum[c]; } if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; } else { return COPY_BLOCK; } } vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride); return FILTER_BLOCK; } int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising) { unsigned char *running_avg_start = running_avg; unsigned char *sig_start = sig; int sum_diff_thresh; int r, c; int sum_diff = 0; int sum_block = 0; int adj_val[3] = { 3, 4, 6 }; int shift_inc1 = 0; int shift_inc2 = 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_UV) { if (increase_denoising) { shift_inc1 = 1; shift_inc2 = 2; } adj_val[0] += shift_inc2; adj_val[1] += shift_inc2; adj_val[2] += shift_inc2; } // Avoid denoising color signal if its close to average level. for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { sum_block += sig[c]; } sig += sig_stride; } if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) { return COPY_BLOCK; } sig -= sig_stride * 8; for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { int diff = 0; int adjustment = 0; int absdiff = 0; diff = mc_running_avg[c] - sig[c]; absdiff = abs(diff); // When |diff| <= |3 + shift_inc1|, use pixel value from // last denoised raw. if (absdiff <= 3 + shift_inc1) { running_avg[c] = mc_running_avg[c]; sum_diff += diff; } else { if (absdiff >= 4 && absdiff <= 7) { adjustment = adj_val[0]; } else if (absdiff >= 8 && absdiff <= 15) { adjustment = adj_val[1]; } else { adjustment = adj_val[2]; } if (diff > 0) { if ((sig[c] + adjustment) > 255) { running_avg[c] = 255; } else { running_avg[c] = sig[c] + adjustment; } sum_diff += adjustment; } else { if ((sig[c] - adjustment) < 0) { running_avg[c] = 0; } else { running_avg[c] = sig[c] - adjustment; } sum_diff -= adjustment; } } } /* Update pointers for next iteration. */ sig += sig_stride; mc_running_avg += mc_avg_stride; running_avg += avg_stride; } sum_diff_thresh = SUM_DIFF_THRESHOLD_UV; if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV; if (abs(sum_diff) > sum_diff_thresh) { // Before returning to copy the block (i.e., apply no denoising), check // if we can still apply some (weaker) temporal filtering to this block, // that would otherwise not be denoised at all. Simplest is to apply // an additional adjustment to running_avg_y to bring it closer to sig. // The adjustment is capped by a maximum delta, and chosen such that // in most cases the resulting sum_diff will be within the // accceptable range given by sum_diff_thresh. // The delta is set by the excess of absolute pixel diff over threshold. int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; // Only apply the adjustment for max delta up to 3. if (delta < 4) { sig -= sig_stride * 8; mc_running_avg -= mc_avg_stride * 8; running_avg -= avg_stride * 8; for (r = 0; r < 8; ++r) { for (c = 0; c < 8; ++c) { int diff = mc_running_avg[c] - sig[c]; int adjustment = abs(diff); if (adjustment > delta) adjustment = delta; if (diff > 0) { // Bring denoised signal down. if (running_avg[c] - adjustment < 0) { running_avg[c] = 0; } else { running_avg[c] = running_avg[c] - adjustment; } sum_diff -= adjustment; } else if (diff < 0) { // Bring denoised signal up. if (running_avg[c] + adjustment > 255) { running_avg[c] = 255; } else { running_avg[c] = running_avg[c] + adjustment; } sum_diff += adjustment; } } // TODO(marpan): Check here if abs(sum_diff) has gone below the // threshold sum_diff_thresh, and if so, we can exit the row loop. sig += sig_stride; mc_running_avg += mc_avg_stride; running_avg += avg_stride; } if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; } else { return COPY_BLOCK; } } vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride); return FILTER_BLOCK; } void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) { assert(mode > 0); // Denoiser is allocated only if mode > 0. if (mode == 1) { denoiser->denoiser_mode = kDenoiserOnYOnly; } else if (mode == 2) { denoiser->denoiser_mode = kDenoiserOnYUV; } else if (mode == 3) { denoiser->denoiser_mode = kDenoiserOnYUVAggressive; } else { denoiser->denoiser_mode = kDenoiserOnYUV; } if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) { denoiser->denoise_pars.scale_sse_thresh = 1; denoiser->denoise_pars.scale_motion_thresh = 8; denoiser->denoise_pars.scale_increase_filter = 0; denoiser->denoise_pars.denoise_mv_bias = 95; denoiser->denoise_pars.pickmode_mv_bias = 100; denoiser->denoise_pars.qp_thresh = 0; denoiser->denoise_pars.consec_zerolast = UINT_MAX; denoiser->denoise_pars.spatial_blur = 0; } else { denoiser->denoise_pars.scale_sse_thresh = 2; denoiser->denoise_pars.scale_motion_thresh = 16; denoiser->denoise_pars.scale_increase_filter = 1; denoiser->denoise_pars.denoise_mv_bias = 60; denoiser->denoise_pars.pickmode_mv_bias = 75; denoiser->denoise_pars.qp_thresh = 80; denoiser->denoise_pars.consec_zerolast = 15; denoiser->denoise_pars.spatial_blur = 0; } } int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height, int num_mb_rows, int num_mb_cols, int mode) { int i; assert(denoiser); denoiser->num_mb_cols = num_mb_cols; for (i = 0; i < MAX_REF_FRAMES; ++i) { denoiser->yv12_running_avg[i].flags = 0; if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_running_avg[i].buffer_alloc, 0, denoiser->yv12_running_avg[i].frame_size); } denoiser->yv12_mc_running_avg.flags = 0; if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0, denoiser->yv12_mc_running_avg.frame_size); if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width, height, VP8BORDERINPIXELS) < 0) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->yv12_last_source.buffer_alloc, 0, denoiser->yv12_last_source.frame_size); denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1); if (!denoiser->denoise_state) { vp8_denoiser_free(denoiser); return 1; } memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols)); vp8_denoiser_set_parameters(denoiser, mode); denoiser->nmse_source_diff = 0; denoiser->nmse_source_diff_count = 0; denoiser->qp_avg = 0; // QP threshold below which we can go up to aggressive mode. denoiser->qp_threshold_up = 80; // QP threshold above which we can go back down to normal mode. // For now keep this second threshold high, so not used currently. denoiser->qp_threshold_down = 128; // Bitrate thresholds and noise metric (nmse) thresholds for switching to // aggressive mode. // TODO(marpan): Adjust thresholds, including effect on resolution. denoiser->bitrate_threshold = 400000; // (bits/sec). denoiser->threshold_aggressive_mode = 80; if (width * height > 1280 * 720) { denoiser->bitrate_threshold = 3000000; denoiser->threshold_aggressive_mode = 200; } else if (width * height > 960 * 540) { denoiser->bitrate_threshold = 1200000; denoiser->threshold_aggressive_mode = 120; } else if (width * height > 640 * 480) { denoiser->bitrate_threshold = 600000; denoiser->threshold_aggressive_mode = 100; } return 0; } void vp8_denoiser_free(VP8_DENOISER *denoiser) { int i; assert(denoiser); for (i = 0; i < MAX_REF_FRAMES; ++i) { vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]); } vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg); vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source); vpx_free(denoiser->denoise_state); } void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser, MACROBLOCK *x, unsigned int best_sse, unsigned int zero_mv_sse, int recon_yoffset, int recon_uvoffset, loop_filter_info_n *lfi_n, int mb_row, int mb_col, int block_index, int consec_zero_last) { int mv_row; int mv_col; unsigned int motion_threshold; unsigned int motion_magnitude2; unsigned int sse_thresh; int sse_diff_thresh = 0; // Spatial loop filter: only applied selectively based on // temporal filter state of block relative to top/left neighbors. int apply_spatial_loop_filter = 1; MV_REFERENCE_FRAME frame = x->best_reference_frame; MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame; enum vp8_denoiser_decision decision = FILTER_BLOCK; enum vp8_denoiser_decision decision_u = COPY_BLOCK; enum vp8_denoiser_decision decision_v = COPY_BLOCK; if (zero_frame) { YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame]; YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg; YV12_BUFFER_CONFIG saved_pre, saved_dst; MB_MODE_INFO saved_mbmi; MACROBLOCKD *filter_xd = &x->e_mbd; MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi; int sse_diff = 0;