ref: 6db96a7f9ca9e569b87720af1c97a231049b2d11
parent: ff36b9c78b8a069978c7ba57c31a8b63fbf8b599
author: Jingning Han <[email protected]>
date: Tue Mar 5 07:14:40 EST 2019
Add normalization over frame level Wiener variance Normalize the block level Wiener variance based decision according to the frame level Wiener variance. Change-Id: Ic2bdf1b322a65661775541dd6c174ba71579461a
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -4732,7 +4732,7 @@
DECLARE_ALIGNED(16, int16_t, src_diff[32 * 32]);
DECLARE_ALIGNED(16, tran_low_t, coeff[32 * 32]);
- int mb_row, mb_col;
+ int mb_row, mb_col, count = 0;
// Hard coded operating block size
const int block_size = 16;
const int coeff_count = block_size * block_size;
@@ -4749,6 +4749,8 @@
memset(zero_pred, 0, sizeof(*zero_pred) * coeff_count);
+ cpi->norm_wiener_variance = 0;
+
for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
int idx, hist_count = 0;
@@ -4789,18 +4791,21 @@
// Wiener filter
for (idx = 1; idx < coeff_count; ++idx) {
- int sign = coeff[idx] < 0;
int64_t sqr_coeff = (int64_t)coeff[idx] * coeff[idx];
coeff[idx] = (int16_t)((sqr_coeff * coeff[idx]) /
(sqr_coeff + (int64_t)median_val * median_val));
- if (sign) coeff[idx] = -coeff[idx];
-
wiener_variance += (int64_t)coeff[idx] * coeff[idx];
}
cpi->mb_wiener_variance[mb_row * cm->mb_cols + mb_col] =
wiener_variance / coeff_count;
+ cpi->norm_wiener_variance +=
+ cpi->mb_wiener_variance[mb_row * cm->mb_cols + mb_col];
+ ++count;
}
}
+
+ if (count) cpi->norm_wiener_variance /= count;
+ cpi->norm_wiener_variance = VPXMAX(1, cpi->norm_wiener_variance);
}
static void encode_frame_to_data_rate(VP9_COMP *cpi, size_t *size,
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -628,6 +628,7 @@
int ext_refresh_frame_context_pending;
int ext_refresh_frame_context;
+ int64_t norm_wiener_variance;
int64_t *mb_wiener_variance;
int *stack_rank_buffer;