Vikas Arora | 1e7bf88 | 2013-03-13 16:43:18 -0700 | [diff] [blame] | 1 | // Copyright 2013 Google Inc. All Rights Reserved. |
| 2 | // |
Vikas Arora | 0406ce1 | 2013-08-09 15:57:12 -0700 | [diff] [blame] | 3 | // Use of this source code is governed by a BSD-style license |
| 4 | // that can be found in the COPYING file in the root of the source |
| 5 | // tree. An additional intellectual property rights grant can be found |
| 6 | // in the file PATENTS. All contributing project authors may |
| 7 | // be found in the AUTHORS file in the root of the source tree. |
Vikas Arora | 1e7bf88 | 2013-03-13 16:43:18 -0700 | [diff] [blame] | 8 | // ----------------------------------------------------------------------------- |
| 9 | // |
Vikas Arora | af51b94 | 2014-08-28 10:51:12 -0700 | [diff] [blame] | 10 | // Implement gradient smoothing: we replace a current alpha value by its |
| 11 | // surrounding average if it's close enough (that is: the change will be less |
| 12 | // than the minimum distance between two quantized level). |
| 13 | // We use sliding window for computing the 2d moving average. |
Vikas Arora | 1e7bf88 | 2013-03-13 16:43:18 -0700 | [diff] [blame] | 14 | // |
| 15 | // Author: Skal (pascal.massimino@gmail.com) |
| 16 | |
| 17 | #include "./quant_levels_dec.h" |
| 18 | |
Vikas Arora | af51b94 | 2014-08-28 10:51:12 -0700 | [diff] [blame] | 19 | #include <string.h> // for memset |
| 20 | |
| 21 | #include "./utils.h" |
| 22 | |
| 23 | // #define USE_DITHERING // uncomment to enable ordered dithering (not vital) |
| 24 | |
| 25 | #define FIX 16 // fix-point precision for averaging |
| 26 | #define LFIX 2 // extra precision for look-up table |
| 27 | #define LUT_SIZE ((1 << (8 + LFIX)) - 1) // look-up table size |
| 28 | |
| 29 | #if defined(USE_DITHERING) |
| 30 | |
| 31 | #define DFIX 4 // extra precision for ordered dithering |
| 32 | #define DSIZE 4 // dithering size (must be a power of two) |
| 33 | // cf. http://en.wikipedia.org/wiki/Ordered_dithering |
| 34 | static const uint8_t kOrderedDither[DSIZE][DSIZE] = { |
| 35 | { 0, 8, 2, 10 }, // coefficients are in DFIX fixed-point precision |
| 36 | { 12, 4, 14, 6 }, |
| 37 | { 3, 11, 1, 9 }, |
| 38 | { 15, 7, 13, 5 } |
| 39 | }; |
| 40 | |
| 41 | #else |
| 42 | #define DFIX 0 |
| 43 | #endif |
| 44 | |
| 45 | typedef struct { |
| 46 | int width_, height_; // dimension |
| 47 | int row_; // current input row being processed |
| 48 | uint8_t* src_; // input pointer |
| 49 | uint8_t* dst_; // output pointer |
| 50 | |
| 51 | int radius_; // filter radius (=delay) |
| 52 | int scale_; // normalization factor, in FIX bits precision |
| 53 | |
| 54 | void* mem_; // all memory |
| 55 | |
| 56 | // various scratch buffers |
| 57 | uint16_t* start_; |
| 58 | uint16_t* cur_; |
| 59 | uint16_t* end_; |
| 60 | uint16_t* top_; |
| 61 | uint16_t* average_; |
| 62 | |
| 63 | // input levels distribution |
| 64 | int num_levels_; // number of quantized levels |
| 65 | int min_, max_; // min and max level values |
| 66 | int min_level_dist_; // smallest distance between two consecutive levels |
| 67 | |
| 68 | int16_t* correction_; // size = 1 + 2*LUT_SIZE -> ~4k memory |
| 69 | } SmoothParams; |
| 70 | |
| 71 | //------------------------------------------------------------------------------ |
| 72 | |
| 73 | #define CLIP_MASK (int)(~0U << (8 + DFIX)) |
| 74 | static WEBP_INLINE uint8_t clip_8b(int v) { |
| 75 | return (!(v & CLIP_MASK)) ? (uint8_t)(v >> DFIX) : (v < 0) ? 0u : 255u; |
| 76 | } |
| 77 | |
| 78 | // vertical accumulation |
| 79 | static void VFilter(SmoothParams* const p) { |
| 80 | const uint8_t* src = p->src_; |
| 81 | const int w = p->width_; |
| 82 | uint16_t* const cur = p->cur_; |
| 83 | const uint16_t* const top = p->top_; |
| 84 | uint16_t* const out = p->end_; |
| 85 | uint16_t sum = 0; // all arithmetic is modulo 16bit |
| 86 | int x; |
| 87 | |
| 88 | for (x = 0; x < w; ++x) { |
| 89 | uint16_t new_value; |
| 90 | sum += src[x]; |
| 91 | new_value = top[x] + sum; |
| 92 | out[x] = new_value - cur[x]; // vertical sum of 'r' pixels. |
| 93 | cur[x] = new_value; |
Vikas Arora | 8b72022 | 2014-01-02 16:48:02 -0800 | [diff] [blame] | 94 | } |
Vikas Arora | af51b94 | 2014-08-28 10:51:12 -0700 | [diff] [blame] | 95 | // move input pointers one row down |
| 96 | p->top_ = p->cur_; |
| 97 | p->cur_ += w; |
| 98 | if (p->cur_ == p->end_) p->cur_ = p->start_; // roll-over |
| 99 | // We replicate edges, as it's somewhat easier as a boundary condition. |
| 100 | // That's why we don't update the 'src' pointer on top/bottom area: |
| 101 | if (p->row_ >= 0 && p->row_ < p->height_ - 1) { |
| 102 | p->src_ += p->width_; |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | // horizontal accumulation. We use mirror replication of missing pixels, as it's |
| 107 | // a little easier to implement (surprisingly). |
| 108 | static void HFilter(SmoothParams* const p) { |
| 109 | const uint16_t* const in = p->end_; |
| 110 | uint16_t* const out = p->average_; |
| 111 | const uint32_t scale = p->scale_; |
| 112 | const int w = p->width_; |
| 113 | const int r = p->radius_; |
| 114 | |
| 115 | int x; |
| 116 | for (x = 0; x <= r; ++x) { // left mirroring |
| 117 | const uint16_t delta = in[x + r - 1] + in[r - x]; |
| 118 | out[x] = (delta * scale) >> FIX; |
| 119 | } |
| 120 | for (; x < w - r; ++x) { // bulk middle run |
| 121 | const uint16_t delta = in[x + r] - in[x - r - 1]; |
| 122 | out[x] = (delta * scale) >> FIX; |
| 123 | } |
| 124 | for (; x < w; ++x) { // right mirroring |
| 125 | const uint16_t delta = |
| 126 | 2 * in[w - 1] - in[2 * w - 2 - r - x] - in[x - r - 1]; |
| 127 | out[x] = (delta * scale) >> FIX; |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | // emit one filtered output row |
| 132 | static void ApplyFilter(SmoothParams* const p) { |
| 133 | const uint16_t* const average = p->average_; |
| 134 | const int w = p->width_; |
| 135 | const int16_t* const correction = p->correction_; |
| 136 | #if defined(USE_DITHERING) |
| 137 | const uint8_t* const dither = kOrderedDither[p->row_ % DSIZE]; |
| 138 | #endif |
| 139 | uint8_t* const dst = p->dst_; |
| 140 | int x; |
| 141 | for (x = 0; x < w; ++x) { |
| 142 | const int v = dst[x]; |
| 143 | if (v < p->max_ && v > p->min_) { |
| 144 | const int c = (v << DFIX) + correction[average[x] - (v << LFIX)]; |
| 145 | #if defined(USE_DITHERING) |
| 146 | dst[x] = clip_8b(c + dither[x % DSIZE]); |
| 147 | #else |
| 148 | dst[x] = clip_8b(c); |
| 149 | #endif |
| 150 | } |
| 151 | } |
| 152 | p->dst_ += w; // advance output pointer |
| 153 | } |
| 154 | |
| 155 | //------------------------------------------------------------------------------ |
| 156 | // Initialize correction table |
| 157 | |
| 158 | static void InitCorrectionLUT(int16_t* const lut, int min_dist) { |
| 159 | // The correction curve is: |
| 160 | // f(x) = x for x <= threshold2 |
| 161 | // f(x) = 0 for x >= threshold1 |
| 162 | // and a linear interpolation for range x=[threshold2, threshold1] |
| 163 | // (along with f(-x) = -f(x) symmetry). |
| 164 | // Note that: threshold2 = 3/4 * threshold1 |
| 165 | const int threshold1 = min_dist << LFIX; |
| 166 | const int threshold2 = (3 * threshold1) >> 2; |
| 167 | const int max_threshold = threshold2 << DFIX; |
| 168 | const int delta = threshold1 - threshold2; |
| 169 | int i; |
| 170 | for (i = 1; i <= LUT_SIZE; ++i) { |
| 171 | int c = (i <= threshold2) ? (i << DFIX) |
| 172 | : (i < threshold1) ? max_threshold * (threshold1 - i) / delta |
| 173 | : 0; |
| 174 | c >>= LFIX; |
| 175 | lut[+i] = +c; |
| 176 | lut[-i] = -c; |
| 177 | } |
| 178 | lut[0] = 0; |
| 179 | } |
| 180 | |
| 181 | static void CountLevels(const uint8_t* const data, int size, |
| 182 | SmoothParams* const p) { |
| 183 | int i, last_level; |
| 184 | uint8_t used_levels[256] = { 0 }; |
| 185 | p->min_ = 255; |
| 186 | p->max_ = 0; |
| 187 | for (i = 0; i < size; ++i) { |
| 188 | const int v = data[i]; |
| 189 | if (v < p->min_) p->min_ = v; |
| 190 | if (v > p->max_) p->max_ = v; |
| 191 | used_levels[v] = 1; |
| 192 | } |
| 193 | // Compute the mininum distance between two non-zero levels. |
| 194 | p->min_level_dist_ = p->max_ - p->min_; |
| 195 | last_level = -1; |
| 196 | for (i = 0; i < 256; ++i) { |
| 197 | if (used_levels[i]) { |
| 198 | ++p->num_levels_; |
| 199 | if (last_level >= 0) { |
| 200 | const int level_dist = i - last_level; |
| 201 | if (level_dist < p->min_level_dist_) { |
| 202 | p->min_level_dist_ = level_dist; |
| 203 | } |
| 204 | } |
| 205 | last_level = i; |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | // Initialize all params. |
| 211 | static int InitParams(uint8_t* const data, int width, int height, |
| 212 | int radius, SmoothParams* const p) { |
| 213 | const int R = 2 * radius + 1; // total size of the kernel |
| 214 | |
| 215 | const size_t size_scratch_m = (R + 1) * width * sizeof(*p->start_); |
| 216 | const size_t size_m = width * sizeof(*p->average_); |
| 217 | const size_t size_lut = (1 + 2 * LUT_SIZE) * sizeof(*p->correction_); |
| 218 | const size_t total_size = size_scratch_m + size_m + size_lut; |
| 219 | uint8_t* mem = (uint8_t*)WebPSafeMalloc(1U, total_size); |
| 220 | |
| 221 | if (mem == NULL) return 0; |
| 222 | p->mem_ = (void*)mem; |
| 223 | |
| 224 | p->start_ = (uint16_t*)mem; |
| 225 | p->cur_ = p->start_; |
| 226 | p->end_ = p->start_ + R * width; |
| 227 | p->top_ = p->end_ - width; |
| 228 | memset(p->top_, 0, width * sizeof(*p->top_)); |
| 229 | mem += size_scratch_m; |
| 230 | |
| 231 | p->average_ = (uint16_t*)mem; |
| 232 | mem += size_m; |
| 233 | |
| 234 | p->width_ = width; |
| 235 | p->height_ = height; |
| 236 | p->src_ = data; |
| 237 | p->dst_ = data; |
| 238 | p->radius_ = radius; |
| 239 | p->scale_ = (1 << (FIX + LFIX)) / (R * R); // normalization constant |
| 240 | p->row_ = -radius; |
| 241 | |
| 242 | // analyze the input distribution so we can best-fit the threshold |
| 243 | CountLevels(data, width * height, p); |
| 244 | |
| 245 | // correction table |
| 246 | p->correction_ = ((int16_t*)mem) + LUT_SIZE; |
| 247 | InitCorrectionLUT(p->correction_, p->min_level_dist_); |
| 248 | |
Vikas Arora | 1e7bf88 | 2013-03-13 16:43:18 -0700 | [diff] [blame] | 249 | return 1; |
| 250 | } |
| 251 | |
Vikas Arora | af51b94 | 2014-08-28 10:51:12 -0700 | [diff] [blame] | 252 | static void CleanupParams(SmoothParams* const p) { |
| 253 | WebPSafeFree(p->mem_); |
| 254 | } |
| 255 | |
| 256 | int WebPDequantizeLevels(uint8_t* const data, int width, int height, |
| 257 | int strength) { |
| 258 | const int radius = 4 * strength / 100; |
| 259 | if (strength < 0 || strength > 100) return 0; |
| 260 | if (data == NULL || width <= 0 || height <= 0) return 0; // bad params |
| 261 | if (radius > 0) { |
| 262 | SmoothParams p; |
| 263 | memset(&p, 0, sizeof(p)); |
| 264 | if (!InitParams(data, width, height, radius, &p)) return 0; |
| 265 | if (p.num_levels_ > 2) { |
| 266 | for (; p.row_ < p.height_; ++p.row_) { |
| 267 | VFilter(&p); // accumulate average of input |
| 268 | // Need to wait few rows in order to prime the filter, |
| 269 | // before emitting some output. |
| 270 | if (p.row_ >= p.radius_) { |
| 271 | HFilter(&p); |
| 272 | ApplyFilter(&p); |
| 273 | } |
| 274 | } |
| 275 | } |
| 276 | CleanupParams(&p); |
| 277 | } |
| 278 | return 1; |
| 279 | } |