Merge "Add an adaptive denoising mode."

3 files changed, 155 insertions, 3 deletions

diff --git a/vp8/encoder/denoising.c b/vp8/encoder/denoising.c
index 75401fc2b..4621d13e2 100644
--- a/vp8/encoder/denoising.c
+++ b/vp8/encoder/denoising.c
@@ -341,8 +341,10 @@ void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) {
     denoiser->denoiser_mode = kDenoiserOnYOnly;
   } else if (mode == 2) {
     denoiser->denoiser_mode = kDenoiserOnYUV;
-  } else {
+  } else if (mode == 3) {
     denoiser->denoiser_mode = kDenoiserOnYUVAggressive;
+  } else {
+    denoiser->denoiser_mode = kDenoiserOnAdaptive;
   }
   if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) {
     denoiser->denoise_pars.scale_sse_thresh = 1;
@@ -397,9 +399,23 @@ int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
     vpx_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;
+    }
+    vpx_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);
     vpx_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;
+    // TODO(marpan): Adjust thresholds, including effect on resolution.
+    denoiser->threshold_aggressive_mode = 40;
+    if (width * height > 640 * 480)
+      denoiser->threshold_aggressive_mode = 180;
     return 0;
 }
 
diff --git a/vp8/encoder/denoising.h b/vp8/encoder/denoising.h
index 89832d3c2..e1844f18a 100644
--- a/vp8/encoder/denoising.h
+++ b/vp8/encoder/denoising.h
@@ -24,7 +24,7 @@ extern "C" {
 
 #define SUM_DIFF_THRESHOLD_UV (96)   // (8 * 8 * 1.5)
 #define SUM_DIFF_THRESHOLD_HIGH_UV (8 * 8 * 2)
-#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 4)
+#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 8)
 #define MOTION_MAGNITUDE_THRESHOLD_UV (8*3)
 
 enum vp8_denoiser_decision
@@ -43,7 +43,8 @@ enum vp8_denoiser_mode {
   kDenoiserOff,
   kDenoiserOnYOnly,
   kDenoiserOnYUV,
-  kDenoiserOnYUVAggressive
+  kDenoiserOnYUVAggressive,
+  kDenoiserOnAdaptive
 };
 
 typedef struct {
@@ -72,9 +73,14 @@ typedef struct vp8_denoiser
 {
     YV12_BUFFER_CONFIG yv12_running_avg[MAX_REF_FRAMES];
     YV12_BUFFER_CONFIG yv12_mc_running_avg;
+    // TODO(marpan): Should remove yv12_last_source and use vp8_lookahead_peak.
+    YV12_BUFFER_CONFIG yv12_last_source;
     unsigned char* denoise_state;
     int num_mb_cols;
     int denoiser_mode;
+    int threshold_aggressive_mode;
+    int nmse_source_diff;
+    int nmse_source_diff_count;
     denoise_params denoise_pars;
 } VP8_DENOISER;
 
@@ -83,6 +89,8 @@ int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
 
 void vp8_denoiser_free(VP8_DENOISER *denoiser);
 
+void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode);
+
 void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                              MACROBLOCK *x,
                              unsigned int best_sse,
diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index f5080b206..93f882507 100644
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -3283,12 +3283,119 @@ static void update_reference_frames(VP8_COMP *cpi)
                         &cpi->denoiser.yv12_running_avg[LAST_FRAME]);
             }
         }
+        if (cpi->oxcf.noise_sensitivity == 4)
+          vp8_yv12_copy_frame(cpi->Source, &cpi->denoiser.yv12_last_source);
 
     }
 #endif
 
 }
 
+static void process_denoiser_mode_change(VP8_COMP *cpi) {
+  const VP8_COMMON *const cm = &cpi->common;
+  int i, j;
+  int total = 0;
+  int num_blocks = 0;
+  // Number of blocks skipped along row/column in computing the
+  // nmse (normalized mean square error) of source.
+  int skip = 2;
+  // Only select blocks for computing nmse that have been encoded
+  // as ZERO LAST min_consec_zero_last frames in a row.
+  int min_consec_zero_last = 10;
+  // Decision is tested for changing the denoising mode every
+  // num_mode_change times this function is called. Note that this
+  // function called every 8 frames, so (8 * num_mode_change) is number
+  // of frames where denoising mode change is tested for switch.
+  int num_mode_change = 15;
+  // Framerate factor, to compensate for larger mse at lower framerates.
+  // TODO(marpan): Adjust this factor,
+  int fac_framerate = cpi->output_framerate < 25.0f ? 80 : 100;
+  int tot_num_blocks = cm->mb_rows * cm->mb_cols;
+  int ystride = cpi->Source->y_stride;
+  unsigned char *src = cpi->Source->y_buffer;
+  unsigned char *dst = cpi->denoiser.yv12_last_source.y_buffer;
+  static const unsigned char const_source[16] = {
+      128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+      128, 128, 128};
+
+  // Loop through the Y plane, every skip blocks along rows and columns,
+  // summing the normalized mean square error, only for blocks that have
+  // been encoded as ZEROMV LAST at least min_consec_zero_last least frames in
+  // a row and have small sum difference between current and previous frame.
+  // Normalization here is by the contrast of the current frame block.
+  for (i = 0; i < cm->Height; i += 16 * skip) {
+    int block_index_row = (i >> 4) * cm->mb_cols;
+    for (j = 0; j < cm->Width; j += 16 * skip) {
+      int index = block_index_row + (j >> 4);
+      if (cpi->consec_zero_last[index] >= min_consec_zero_last) {
+        unsigned int sse;
+        const unsigned int mse = vp8_mse16x16(src + j,
+                                              ystride,
+                                              dst + j,
+                                              ystride,
+                                              &sse);
+        const unsigned int var = vp8_variance16x16(src + j,
+                                                   ystride,
+                                                   dst + j,
+                                                   ystride,
+                                                   &sse);
+        // Only consider this block as valid for noise measurement
+        // if the sum_diff average of the current and previous frame
+        // is small (to avoid effects from lighting change).
+        if ((mse - var) < 256) {
+          const unsigned int act = vp8_variance16x16(src + j,
+                                                     ystride,
+                                                     const_source,
+                                                     0,
+                                                     &sse);
+          if (act > 0)
+            total += mse / act;
+          num_blocks++;
+        }
+      }
+    }
+    src += 16 * skip * ystride;
+    dst += 16 * skip * ystride;
+  }
+  total = total * fac_framerate / 100;
+
+  // Only consider this frame as valid sample if we have computed nmse over
+  // at least ~1/16 blocks, and Total > 0 (Total == 0 can happen if the
+  // application inputs duplicate frames, or contrast is all zero).
+  if (total > 0 &&
+      (num_blocks > (tot_num_blocks >> 4))) {
+    // Update the recursive mean square source_diff.
+    if (cpi->denoiser.nmse_source_diff_count == 0)
+      // First sample in new interval.
+      cpi->denoiser.nmse_source_diff = total;
+    else
+      // For subsequent samples, use average with weight ~1/4 for new sample.
+      cpi->denoiser.nmse_source_diff = (int)((total >> 2) +
+          3 * (cpi->denoiser.nmse_source_diff >> 2));
+    cpi->denoiser.nmse_source_diff_count++;
+  }
+  // Check for changing the denoiser mode, when we have obtained #samples =
+  // num_mode_change.
+  if (cpi->denoiser.nmse_source_diff_count == num_mode_change) {
+    // Check for going up: from normal to aggressive mode.
+    if ((cpi->denoiser.denoiser_mode = kDenoiserOnYUV) &&
+        (cpi->denoiser.nmse_source_diff >
+        cpi->denoiser.threshold_aggressive_mode)) {
+      vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUVAggressive);
+    } else {
+      // Check for going down: from aggressive to normal mode.
+      if ((cpi->denoiser.denoiser_mode = kDenoiserOnYUVAggressive) &&
+          (cpi->denoiser.nmse_source_diff <
+          cpi->denoiser.threshold_aggressive_mode)) {
+        vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUV);
+      }
+    }
+    // Reset metric and counter for next interval.
+    cpi->denoiser.nmse_source_diff = 0;
+    cpi->denoiser.nmse_source_diff_count = 0;
+  }
+}
+
 void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm)
 {
     const FRAME_TYPE frame_type = cm->frame_type;
@@ -3445,6 +3552,12 @@ static void encode_frame_to_data_rate
     {
         /* Key frame from VFW/auto-keyframe/first frame */
         cm->frame_type = KEY_FRAME;
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity == 4) {
+          // For adaptive mode, reset denoiser to normal mode on key frame.
+          vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUV);
+        }
+#endif
     }
 
 #if CONFIG_MULTI_RES_ENCODING
@@ -4472,6 +4585,21 @@ static void encode_frame_to_data_rate
 
     cm->frame_to_show = &cm->yv12_fb[cm->new_fb_idx];
 
+#if CONFIG_TEMPORAL_DENOISING
+    // For the adaptive denoising mode (noise_sensitivity == 4), sample the mse
+    // of source diff (between current and previous frame), and determine if we
+    // should switch the denoiser mode. Sampling refers to computing the mse for
+    // a sub-sample of the frame (i.e., skip x blocks along row/column), and
+    // only for blocks in that set that have used ZEROMV LAST, along with some
+    // constraint on the sum diff between blocks. This process is called every
+    // ~8 frames, to further reduce complexity.
+    if (cpi->oxcf.noise_sensitivity == 4 &&
+        cpi->frames_since_key % 8 == 0 &&
+        cm->frame_type != KEY_FRAME) {
+      process_denoiser_mode_change(cpi);
+    }
+#endif
+
 #if CONFIG_MULTITHREAD
     if (cpi->b_multi_threaded)
     {