2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
10 * This code was originally written by: Gregory Maxwell, at the Daala
17 #include "./vpx_config.h"
18 #include "./vp9_rtcd.h"
19 #include "./vpx_dsp_rtcd.h"
20 #include "vp9/encoder/vp9_ssim.h"
23 # define M_PI (3.141592653589793238462643)
27 void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, int xstride) {
29 vpx_fdct8x8(x, y, ystride);
32 /* Normalized inverse quantization matrix for 8x8 DCT at the point of
33 * transparency. This is not the JPEG based matrix from the paper,
34 this one gives a slightly higher MOS agreement.*/
35 float csf_y[8][8] = {{1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411,
36 1.00227514334, 0.678296995242, 0.466224900598, 0.3265091542}, {2.2901594831,
37 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963, 0.868920337363,
38 0.61280991668, 0.436405793551}, {2.08509755623, 2.04793073064,
39 1.34329019223, 1.09205635862, 0.875748795257, 0.670882927016,
40 0.501731932449, 0.372504254596}, {1.48366094411, 1.68731108984,
41 1.09205635862, 0.772819797575, 0.605636379554, 0.48309405692,
42 0.380429446972, 0.295774038565}, {1.00227514334, 1.2305666963,
43 0.875748795257, 0.605636379554, 0.448996256676, 0.352889268808,
44 0.283006984131, 0.226951348204}, {0.678296995242, 0.868920337363,
45 0.670882927016, 0.48309405692, 0.352889268808, 0.27032073436,
46 0.215017739696, 0.17408067321}, {0.466224900598, 0.61280991668,
47 0.501731932449, 0.380429446972, 0.283006984131, 0.215017739696,
48 0.168869545842, 0.136153931001}, {0.3265091542, 0.436405793551,
49 0.372504254596, 0.295774038565, 0.226951348204, 0.17408067321,
50 0.136153931001, 0.109083846276}};
51 float csf_cb420[8][8] = {
52 {1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788,
53 0.898018824055, 0.74725392039, 0.615105596242}, {2.46074210438,
54 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972,
55 1.17428548929, 0.996404342439, 0.830890433625}, {1.18284184739,
56 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362,
57 0.960060382087, 0.849823426169, 0.731221236837}, {1.14982565193,
58 1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099,
59 0.751437590932, 0.685398513368, 0.608694761374}, {1.05017074788,
60 1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187,
61 0.605503172737, 0.55002013668, 0.495804539034}, {0.898018824055,
62 1.17428548929, 0.960060382087, 0.751437590932, 0.605503172737,
63 0.514674450957, 0.454353482512, 0.407050308965}, {0.74725392039,
64 0.996404342439, 0.849823426169, 0.685398513368, 0.55002013668,
65 0.454353482512, 0.389234902883, 0.342353999733}, {0.615105596242,
66 0.830890433625, 0.731221236837, 0.608694761374, 0.495804539034,
67 0.407050308965, 0.342353999733, 0.295530605237}};
68 float csf_cr420[8][8] = {
69 {2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469,
70 0.867069376285, 0.721500455585, 0.593906509971}, {2.62502345193,
71 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198,
72 1.13381474809, 0.962064122248, 0.802254508198}, {1.26180942886,
73 1.17180569821, 0.944981930573, 0.990876405848, 0.995903384143,
74 0.926972725286, 0.820534991409, 0.706020324706}, {1.11019789803,
75 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195,
76 0.725539939514, 0.661776842059, 0.587716619023}, {1.01397751469,
77 1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286,
78 0.584635025748, 0.531064164893, 0.478717061273}, {0.867069376285,
79 1.13381474809, 0.926972725286, 0.725539939514, 0.584635025748,
80 0.496936637883, 0.438694579826, 0.393021669543}, {0.721500455585,
81 0.962064122248, 0.820534991409, 0.661776842059, 0.531064164893,
82 0.438694579826, 0.375820256136, 0.330555063063}, {0.593906509971,
83 0.802254508198, 0.706020324706, 0.587716619023, 0.478717061273,
84 0.393021669543, 0.330555063063, 0.285345396658}};
86 static double convert_score_db(double _score, double _weight) {
87 return 10 * (log10(255 * 255) - log10(_weight * _score));
90 static double calc_psnrhvs(const unsigned char *_src, int _systride,
91 const unsigned char *_dst, int _dystride,
92 double _par, int _w, int _h, int _step,
95 int16_t dct_s[8 * 8], dct_d[8 * 8];
96 tran_low_t dct_s_coef[8 * 8], dct_d_coef[8 * 8];
103 /*In the PSNR-HVS-M paper[1] the authors describe the construction of
104 their masking table as "we have used the quantization table for the
105 color component Y of JPEG [6] that has been also obtained on the
106 basis of CSF. Note that the values in quantization table JPEG have
107 been normalized and then squared." Their CSF matrix (from PSNR-HVS)
108 was also constructed from the JPEG matrices. I can not find any obvious
109 scheme of normalizing to produce their table, but if I multiply their
110 CSF by 0.38857 and square the result I get their masking table.
111 I have no idea where this constant comes from, but deviating from it
112 too greatly hurts MOS agreement.
114 [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli,
115 Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking
116 of DCT basis functions", CD-ROM Proceedings of the Third
117 International Workshop on Video Processing and Quality Metrics for Consumer
118 Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/
119 for (x = 0; x < 8; x++)
120 for (y = 0; y < 8; y++)
121 mask[x][y] = (_csf[x][y] * 0.3885746225901003)
122 * (_csf[x][y] * 0.3885746225901003);
123 for (y = 0; y < _h - 7; y += _step) {
124 for (x = 0; x < _w - 7; x += _step) {
137 for (i = 0; i < 4; i++)
138 s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0;
139 for (i = 0; i < 8; i++) {
140 for (j = 0; j < 8; j++) {
141 int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
142 dct_s[i * 8 + j] = _src[(y + i) * _systride + (j + x)];
143 dct_d[i * 8 + j] = _dst[(y + i) * _dystride + (j + x)];
144 s_gmean += dct_s[i * 8 + j];
145 d_gmean += dct_d[i * 8 + j];
146 s_means[sub] += dct_s[i * 8 + j];
147 d_means[sub] += dct_d[i * 8 + j];
152 for (i = 0; i < 4; i++)
154 for (i = 0; i < 4; i++)
156 for (i = 0; i < 8; i++) {
157 for (j = 0; j < 8; j++) {
158 int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
159 s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean);
160 d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean);
161 s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub])
162 * (dct_s[i * 8 + j] - s_means[sub]);
163 d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub])
164 * (dct_d[i * 8 + j] - d_means[sub]);
167 s_gvar *= 1 / 63.f * 64;
168 d_gvar *= 1 / 63.f * 64;
169 for (i = 0; i < 4; i++)
170 s_vars[i] *= 1 / 15.f * 16;
171 for (i = 0; i < 4; i++)
172 d_vars[i] *= 1 / 15.f * 16;
174 s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar;
176 d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar;
177 od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8);
178 od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8);
179 for (i = 0; i < 8; i++)
180 for (j = (i == 0); j < 8; j++)
181 s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j];
182 for (i = 0; i < 8; i++)
183 for (j = (i == 0); j < 8; j++)
184 d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j];
185 s_mask = sqrt(s_mask * s_gvar) / 32.f;
186 d_mask = sqrt(d_mask * d_gvar) / 32.f;
189 for (i = 0; i < 8; i++) {
190 for (j = 0; j < 8; j++) {
192 err = fabs(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j]);
193 if (i != 0 || j != 0)
194 err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j];
195 ret += (err * _csf[i][j]) * (err * _csf[i][j]);
204 double vp9_psnrhvs(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
205 double *y_psnrhvs, double *u_psnrhvs, double *v_psnrhvs) {
209 vp9_clear_system_state();
210 *y_psnrhvs = calc_psnrhvs(source->y_buffer, source->y_stride, dest->y_buffer,
211 dest->y_stride, par, source->y_crop_width,
212 source->y_crop_height, step, csf_y);
214 *u_psnrhvs = calc_psnrhvs(source->u_buffer, source->uv_stride, dest->u_buffer,
215 dest->uv_stride, par, source->uv_crop_width,
216 source->uv_crop_height, step, csf_cb420);
218 *v_psnrhvs = calc_psnrhvs(source->v_buffer, source->uv_stride, dest->v_buffer,
219 dest->uv_stride, par, source->uv_crop_width,
220 source->uv_crop_height, step, csf_cr420);
221 psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs));
223 return convert_score_db(psnrhvs, 1.0);