tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

ssim.c (17419B)


      1 /*
      2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
      3 *
      4 * This source code is subject to the terms of the BSD 2 Clause License and
      5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
      6 * was not distributed with this source code in the LICENSE file, you can
      7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
      8 * Media Patent License 1.0 was not distributed with this source code in the
      9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
     10 */
     11 
     12 #include <assert.h>
     13 #include <math.h>
     14 
     15 #include "config/aom_dsp_rtcd.h"
     16 
     17 #include "aom_dsp/ssim.h"
     18 #include "aom_ports/mem.h"
     19 
     20 void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp,
     21                          uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s,
     22                          uint32_t *sum_sq_r, uint32_t *sum_sxr) {
     23  int i, j;
     24  for (i = 0; i < 8; i++, s += sp, r += rp) {
     25    for (j = 0; j < 8; j++) {
     26      *sum_s += s[j];
     27      *sum_r += r[j];
     28      *sum_sq_s += s[j] * s[j];
     29      *sum_sq_r += r[j] * r[j];
     30      *sum_sxr += s[j] * r[j];
     31    }
     32  }
     33 }
     34 
     35 static const int64_t cc1 = 26634;        // (64^2*(.01*255)^2
     36 static const int64_t cc2 = 239708;       // (64^2*(.03*255)^2
     37 static const int64_t cc1_10 = 428658;    // (64^2*(.01*1023)^2
     38 static const int64_t cc2_10 = 3857925;   // (64^2*(.03*1023)^2
     39 static const int64_t cc1_12 = 6868593;   // (64^2*(.01*4095)^2
     40 static const int64_t cc2_12 = 61817334;  // (64^2*(.03*4095)^2
     41 
     42 static double similarity(uint32_t sum_s, uint32_t sum_r, uint32_t sum_sq_s,
     43                         uint32_t sum_sq_r, uint32_t sum_sxr, int count,
     44                         uint32_t bd) {
     45  double ssim_n, ssim_d;
     46  int64_t c1 = 0, c2 = 0;
     47  if (bd == 8) {
     48    // scale the constants by number of pixels
     49    c1 = (cc1 * count * count) >> 12;
     50    c2 = (cc2 * count * count) >> 12;
     51  } else if (bd == 10) {
     52    c1 = (cc1_10 * count * count) >> 12;
     53    c2 = (cc2_10 * count * count) >> 12;
     54  } else if (bd == 12) {
     55    c1 = (cc1_12 * count * count) >> 12;
     56    c2 = (cc2_12 * count * count) >> 12;
     57  } else {
     58    assert(0);
     59    // Return similarity as zero for unsupported bit-depth values.
     60    return 0;
     61  }
     62 
     63  ssim_n = (2.0 * sum_s * sum_r + c1) *
     64           (2.0 * count * sum_sxr - 2.0 * sum_s * sum_r + c2);
     65 
     66  ssim_d = ((double)sum_s * sum_s + (double)sum_r * sum_r + c1) *
     67           ((double)count * sum_sq_s - (double)sum_s * sum_s +
     68            (double)count * sum_sq_r - (double)sum_r * sum_r + c2);
     69 
     70  return ssim_n / ssim_d;
     71 }
     72 
     73 static double ssim_8x8(const uint8_t *s, int sp, const uint8_t *r, int rp) {
     74  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
     75  aom_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
     76                     &sum_sxr);
     77  return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64, 8);
     78 }
     79 
     80 // We are using a 8x8 moving window with starting location of each 8x8 window
     81 // on the 4x4 pixel grid. Such arrangement allows the windows to overlap
     82 // block boundaries to penalize blocking artifacts.
     83 double aom_ssim2(const uint8_t *img1, const uint8_t *img2, int stride_img1,
     84                 int stride_img2, int width, int height) {
     85  int i, j;
     86  int samples = 0;
     87  double ssim_total = 0;
     88 
     89  // sample point start with each 4x4 location
     90  for (i = 0; i <= height - 8;
     91       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
     92    for (j = 0; j <= width - 8; j += 4) {
     93      double v = ssim_8x8(img1 + j, stride_img1, img2 + j, stride_img2);
     94      ssim_total += v;
     95      samples++;
     96    }
     97  }
     98  ssim_total /= samples;
     99  return ssim_total;
    100 }
    101 
    102 #if CONFIG_INTERNAL_STATS
    103 void aom_lowbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
    104                         const YV12_BUFFER_CONFIG *dest, double *weight,
    105                         double *fast_ssim) {
    106  double abc[3];
    107  for (int i = 0; i < 3; ++i) {
    108    const int is_uv = i > 0;
    109    abc[i] = aom_ssim2(source->buffers[i], dest->buffers[i],
    110                       source->strides[is_uv], dest->strides[is_uv],
    111                       source->crop_widths[is_uv], source->crop_heights[is_uv]);
    112  }
    113 
    114  *weight = 1;
    115  *fast_ssim = abc[0] * .8 + .1 * (abc[1] + abc[2]);
    116 }
    117 
    118 // traditional ssim as per: http://en.wikipedia.org/wiki/Structural_similarity
    119 //
    120 // Re working out the math ->
    121 //
    122 // ssim(x,y) =  (2*mean(x)*mean(y) + c1)*(2*cov(x,y)+c2) /
    123 //   ((mean(x)^2+mean(y)^2+c1)*(var(x)+var(y)+c2))
    124 //
    125 // mean(x) = sum(x) / n
    126 //
    127 // cov(x,y) = (n*sum(xi*yi)-sum(x)*sum(y))/(n*n)
    128 //
    129 // var(x) = (n*sum(xi*xi)-sum(xi)*sum(xi))/(n*n)
    130 //
    131 // ssim(x,y) =
    132 //   (2*sum(x)*sum(y)/(n*n) + c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))/(n*n)+c2) /
    133 //   (((sum(x)*sum(x)+sum(y)*sum(y))/(n*n) +c1) *
    134 //    ((n*sum(xi*xi) - sum(xi)*sum(xi))/(n*n)+
    135 //     (n*sum(yi*yi) - sum(yi)*sum(yi))/(n*n)+c2)))
    136 //
    137 // factoring out n*n
    138 //
    139 // ssim(x,y) =
    140 //   (2*sum(x)*sum(y) + n*n*c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))+n*n*c2) /
    141 //   (((sum(x)*sum(x)+sum(y)*sum(y)) + n*n*c1) *
    142 //    (n*sum(xi*xi)-sum(xi)*sum(xi)+n*sum(yi*yi)-sum(yi)*sum(yi)+n*n*c2))
    143 //
    144 // Replace c1 with n*n * c1 for the final step that leads to this code:
    145 // The final step scales by 12 bits so we don't lose precision in the constants.
    146 
    147 static double ssimv_similarity(const Ssimv *sv, int64_t n) {
    148  // Scale the constants by number of pixels.
    149  const int64_t c1 = (cc1 * n * n) >> 12;
    150  const int64_t c2 = (cc2 * n * n) >> 12;
    151 
    152  const double l = 1.0 * (2 * sv->sum_s * sv->sum_r + c1) /
    153                   (sv->sum_s * sv->sum_s + sv->sum_r * sv->sum_r + c1);
    154 
    155  // Since these variables are unsigned sums, convert to double so
    156  // math is done in double arithmetic.
    157  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) /
    158                   (n * sv->sum_sq_s - sv->sum_s * sv->sum_s +
    159                    n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2);
    160 
    161  return l * v;
    162 }
    163 
    164 // The first term of the ssim metric is a luminance factor.
    165 //
    166 // (2*mean(x)*mean(y) + c1)/ (mean(x)^2+mean(y)^2+c1)
    167 //
    168 // This luminance factor is super sensitive to the dark side of luminance
    169 // values and completely insensitive on the white side.  check out 2 sets
    170 // (1,3) and (250,252) the term gives ( 2*1*3/(1+9) = .60
    171 // 2*250*252/ (250^2+252^2) => .99999997
    172 //
    173 // As a result in this tweaked version of the calculation in which the
    174 // luminance is taken as percentage off from peak possible.
    175 //
    176 // 255 * 255 - (sum_s - sum_r) / count * (sum_s - sum_r) / count
    177 //
    178 static double ssimv_similarity2(const Ssimv *sv, int64_t n) {
    179  // Scale the constants by number of pixels.
    180  const int64_t c1 = (cc1 * n * n) >> 12;
    181  const int64_t c2 = (cc2 * n * n) >> 12;
    182 
    183  const double mean_diff = (1.0 * sv->sum_s - sv->sum_r) / n;
    184  const double l = (255 * 255 - mean_diff * mean_diff + c1) / (255 * 255 + c1);
    185 
    186  // Since these variables are unsigned, sums convert to double so
    187  // math is done in double arithmetic.
    188  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) /
    189                   (n * sv->sum_sq_s - sv->sum_s * sv->sum_s +
    190                    n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2);
    191 
    192  return l * v;
    193 }
    194 static void ssimv_parms(uint8_t *img1, int img1_pitch, uint8_t *img2,
    195                        int img2_pitch, Ssimv *sv) {
    196  aom_ssim_parms_8x8(img1, img1_pitch, img2, img2_pitch, &sv->sum_s, &sv->sum_r,
    197                     &sv->sum_sq_s, &sv->sum_sq_r, &sv->sum_sxr);
    198 }
    199 
    200 double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
    201                            int img2_pitch, int width, int height, Ssimv *sv2,
    202                            Metrics *m, int do_inconsistency) {
    203  double dssim_total = 0;
    204  double ssim_total = 0;
    205  double ssim2_total = 0;
    206  double inconsistency_total = 0;
    207  int i, j;
    208  int c = 0;
    209  double norm;
    210  double old_ssim_total = 0;
    211  // We can sample points as frequently as we like start with 1 per 4x4.
    212  for (i = 0; i < height;
    213       i += 4, img1 += img1_pitch * 4, img2 += img2_pitch * 4) {
    214    for (j = 0; j < width; j += 4, ++c) {
    215      Ssimv sv = { 0, 0, 0, 0, 0, 0 };
    216      double ssim;
    217      double ssim2;
    218      double dssim;
    219      uint32_t var_new;
    220      uint32_t var_old;
    221      uint32_t mean_new;
    222      uint32_t mean_old;
    223      double ssim_new;
    224      double ssim_old;
    225 
    226      // Not sure there's a great way to handle the edge pixels
    227      // in ssim when using a window. Seems biased against edge pixels
    228      // however you handle this. This uses only samples that are
    229      // fully in the frame.
    230      if (j + 8 <= width && i + 8 <= height) {
    231        ssimv_parms(img1 + j, img1_pitch, img2 + j, img2_pitch, &sv);
    232      }
    233 
    234      ssim = ssimv_similarity(&sv, 64);
    235      ssim2 = ssimv_similarity2(&sv, 64);
    236 
    237      sv.ssim = ssim2;
    238 
    239      // dssim is calculated to use as an actual error metric and
    240      // is scaled up to the same range as sum square error.
    241      // Since we are subsampling every 16th point maybe this should be
    242      // *16 ?
    243      dssim = 255 * 255 * (1 - ssim2) / 2;
    244 
    245      // Here I introduce a new error metric: consistency-weighted
    246      // SSIM-inconsistency.  This metric isolates frames where the
    247      // SSIM 'suddenly' changes, e.g. if one frame in every 8 is much
    248      // sharper or blurrier than the others. Higher values indicate a
    249      // temporally inconsistent SSIM. There are two ideas at work:
    250      //
    251      // 1) 'SSIM-inconsistency': the total inconsistency value
    252      // reflects how much SSIM values are changing between this
    253      // source / reference frame pair and the previous pair.
    254      //
    255      // 2) 'consistency-weighted': weights de-emphasize areas in the
    256      // frame where the scene content has changed. Changes in scene
    257      // content are detected via changes in local variance and local
    258      // mean.
    259      //
    260      // Thus the overall measure reflects how inconsistent the SSIM
    261      // values are, over consistent regions of the frame.
    262      //
    263      // The metric has three terms:
    264      //
    265      // term 1 -> uses change in scene Variance to weight error score
    266      //  2 * var(Fi)*var(Fi-1) / (var(Fi)^2+var(Fi-1)^2)
    267      //  larger changes from one frame to the next mean we care
    268      //  less about consistency.
    269      //
    270      // term 2 -> uses change in local scene luminance to weight error
    271      //  2 * avg(Fi)*avg(Fi-1) / (avg(Fi)^2+avg(Fi-1)^2)
    272      //  larger changes from one frame to the next mean we care
    273      //  less about consistency.
    274      //
    275      // term3 -> measures inconsistency in ssim scores between frames
    276      //   1 - ( 2 * ssim(Fi)*ssim(Fi-1)/(ssim(Fi)^2+sssim(Fi-1)^2).
    277      //
    278      // This term compares the ssim score for the same location in 2
    279      // subsequent frames.
    280      var_new = sv.sum_sq_s - sv.sum_s * sv.sum_s / 64;
    281      var_old = sv2[c].sum_sq_s - sv2[c].sum_s * sv2[c].sum_s / 64;
    282      mean_new = sv.sum_s;
    283      mean_old = sv2[c].sum_s;
    284      ssim_new = sv.ssim;
    285      ssim_old = sv2[c].ssim;
    286 
    287      if (do_inconsistency) {
    288        // We do the metric once for every 4x4 block in the image. Since
    289        // we are scaling the error to SSE for use in a psnr calculation
    290        // 1.0 = 4x4x255x255 the worst error we can possibly have.
    291        static const double kScaling = 4. * 4 * 255 * 255;
    292 
    293        // The constants have to be non 0 to avoid potential divide by 0
    294        // issues other than that they affect kind of a weighting between
    295        // the terms.  No testing of what the right terms should be has been
    296        // done.
    297        static const double c1 = 1, c2 = 1, c3 = 1;
    298 
    299        // This measures how much consistent variance is in two consecutive
    300        // source frames. 1.0 means they have exactly the same variance.
    301        const double variance_term =
    302            (2.0 * var_old * var_new + c1) /
    303            (1.0 * var_old * var_old + 1.0 * var_new * var_new + c1);
    304 
    305        // This measures how consistent the local mean are between two
    306        // consecutive frames. 1.0 means they have exactly the same mean.
    307        const double mean_term =
    308            (2.0 * mean_old * mean_new + c2) /
    309            (1.0 * mean_old * mean_old + 1.0 * mean_new * mean_new + c2);
    310 
    311        // This measures how consistent the ssims of two
    312        // consecutive frames is. 1.0 means they are exactly the same.
    313        double ssim_term =
    314            pow((2.0 * ssim_old * ssim_new + c3) /
    315                    (ssim_old * ssim_old + ssim_new * ssim_new + c3),
    316                5);
    317 
    318        double this_inconsistency;
    319 
    320        // Floating point math sometimes makes this > 1 by a tiny bit.
    321        // We want the metric to scale between 0 and 1.0 so we can convert
    322        // it to an snr scaled value.
    323        if (ssim_term > 1) ssim_term = 1;
    324 
    325        // This converts the consistency metric to an inconsistency metric
    326        // ( so we can scale it like psnr to something like sum square error.
    327        // The reason for the variance and mean terms is the assumption that
    328        // if there are big changes in the source we shouldn't penalize
    329        // inconsistency in ssim scores a bit less as it will be less visible
    330        // to the user.
    331        this_inconsistency = (1 - ssim_term) * variance_term * mean_term;
    332 
    333        this_inconsistency *= kScaling;
    334        inconsistency_total += this_inconsistency;
    335      }
    336      sv2[c] = sv;
    337      ssim_total += ssim;
    338      ssim2_total += ssim2;
    339      dssim_total += dssim;
    340 
    341      old_ssim_total += ssim_old;
    342    }
    343    old_ssim_total += 0;
    344  }
    345 
    346  norm = 1. / (width / 4) / (height / 4);
    347  ssim_total *= norm;
    348  ssim2_total *= norm;
    349  m->ssim2 = ssim2_total;
    350  m->ssim = ssim_total;
    351  if (old_ssim_total == 0) inconsistency_total = 0;
    352 
    353  m->ssimc = inconsistency_total;
    354 
    355  m->dssim = dssim_total;
    356  return inconsistency_total;
    357 }
    358 #endif  // CONFIG_INTERNAL_STATS
    359 
    360 #if CONFIG_AV1_HIGHBITDEPTH
    361 void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r,
    362                                 int rp, uint32_t *sum_s, uint32_t *sum_r,
    363                                 uint32_t *sum_sq_s, uint32_t *sum_sq_r,
    364                                 uint32_t *sum_sxr) {
    365  int i, j;
    366  for (i = 0; i < 8; i++, s += sp, r += rp) {
    367    for (j = 0; j < 8; j++) {
    368      *sum_s += s[j];
    369      *sum_r += r[j];
    370      *sum_sq_s += s[j] * s[j];
    371      *sum_sq_r += r[j] * r[j];
    372      *sum_sxr += s[j] * r[j];
    373    }
    374  }
    375 }
    376 
    377 static double highbd_ssim_8x8(const uint16_t *s, int sp, const uint16_t *r,
    378                              int rp, uint32_t bd, uint32_t shift) {
    379  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
    380  aom_highbd_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
    381                            &sum_sxr);
    382  return similarity(sum_s >> shift, sum_r >> shift, sum_sq_s >> (2 * shift),
    383                    sum_sq_r >> (2 * shift), sum_sxr >> (2 * shift), 64, bd);
    384 }
    385 
    386 double aom_highbd_ssim2(const uint8_t *img1, const uint8_t *img2,
    387                        int stride_img1, int stride_img2, int width, int height,
    388                        uint32_t bd, uint32_t shift) {
    389  int i, j;
    390  int samples = 0;
    391  double ssim_total = 0;
    392 
    393  // sample point start with each 4x4 location
    394  for (i = 0; i <= height - 8;
    395       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
    396    for (j = 0; j <= width - 8; j += 4) {
    397      double v = highbd_ssim_8x8(CONVERT_TO_SHORTPTR(img1 + j), stride_img1,
    398                                 CONVERT_TO_SHORTPTR(img2 + j), stride_img2, bd,
    399                                 shift);
    400      ssim_total += v;
    401      samples++;
    402    }
    403  }
    404  ssim_total /= samples;
    405  return ssim_total;
    406 }
    407 
    408 #if CONFIG_INTERNAL_STATS
    409 void aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
    410                          const YV12_BUFFER_CONFIG *dest, double *weight,
    411                          uint32_t bd, uint32_t in_bd, double *fast_ssim) {
    412  assert(bd >= in_bd);
    413  uint32_t shift = bd - in_bd;
    414 
    415  double abc[3];
    416  for (int i = 0; i < 3; ++i) {
    417    const int is_uv = i > 0;
    418    abc[i] = aom_highbd_ssim2(source->buffers[i], dest->buffers[i],
    419                              source->strides[is_uv], dest->strides[is_uv],
    420                              source->crop_widths[is_uv],
    421                              source->crop_heights[is_uv], in_bd, shift);
    422  }
    423 
    424  weight[0] = 1;
    425  fast_ssim[0] = abc[0] * .8 + .1 * (abc[1] + abc[2]);
    426 
    427  if (bd > in_bd) {
    428    // Compute SSIM based on stream bit depth
    429    shift = 0;
    430    for (int i = 0; i < 3; ++i) {
    431      const int is_uv = i > 0;
    432      abc[i] = aom_highbd_ssim2(source->buffers[i], dest->buffers[i],
    433                                source->strides[is_uv], dest->strides[is_uv],
    434                                source->crop_widths[is_uv],
    435                                source->crop_heights[is_uv], bd, shift);
    436    }
    437 
    438    weight[1] = 1;
    439    fast_ssim[1] = abc[0] * .8 + .1 * (abc[1] + abc[2]);
    440  }
    441 }
    442 #endif  // CONFIG_INTERNAL_STATS
    443 #endif  // CONFIG_AV1_HIGHBITDEPTH
    444 
    445 #if CONFIG_INTERNAL_STATS
    446 void aom_calc_ssim(const YV12_BUFFER_CONFIG *orig,
    447                   const YV12_BUFFER_CONFIG *recon, const uint32_t bit_depth,
    448                   const uint32_t in_bit_depth, int is_hbd, double *weight,
    449                   double *frame_ssim2) {
    450 #if CONFIG_AV1_HIGHBITDEPTH
    451  if (is_hbd) {
    452    aom_highbd_calc_ssim(orig, recon, weight, bit_depth, in_bit_depth,
    453                         frame_ssim2);
    454    return;
    455  }
    456 #else
    457  (void)bit_depth;
    458  (void)in_bit_depth;
    459  (void)is_hbd;
    460 #endif  // CONFIG_AV1_HIGHBITDEPTH
    461  aom_lowbd_calc_ssim(orig, recon, weight, frame_ssim2);
    462 }
    463 #endif  // CONFIG_INTERNAL_STATS