1 | // Copyright 2013 Google Inc. All Rights Reserved. |
2 | // |
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. |
8 | // ----------------------------------------------------------------------------- |
9 | // |
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. |
14 | // |
15 | // Author: Skal (pascal.massimino@gmail.com) |
16 | |
17 | #include "src/utils/quant_levels_dec_utils.h" |
18 | |
19 | #include <string.h> // for memset |
20 | |
21 | #include "src/utils/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. https://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 stride_; // stride in bytes |
48 | int row_; // current input row being processed |
49 | uint8_t* src_; // input pointer |
50 | uint8_t* dst_; // output pointer |
51 | |
52 | int radius_; // filter radius (=delay) |
53 | int scale_; // normalization factor, in FIX bits precision |
54 | |
55 | void* mem_; // all memory |
56 | |
57 | // various scratch buffers |
58 | uint16_t* start_; |
59 | uint16_t* cur_; |
60 | uint16_t* end_; |
61 | uint16_t* top_; |
62 | uint16_t* average_; |
63 | |
64 | // input levels distribution |
65 | int num_levels_; // number of quantized levels |
66 | int min_, max_; // min and max level values |
67 | int min_level_dist_; // smallest distance between two consecutive levels |
68 | |
69 | int16_t* correction_; // size = 1 + 2*LUT_SIZE -> ~4k memory |
70 | } SmoothParams; |
71 | |
72 | //------------------------------------------------------------------------------ |
73 | |
74 | #define CLIP_8b_MASK (int)(~0U << (8 + DFIX)) |
75 | static WEBP_INLINE uint8_t clip_8b(int v) { |
76 | return (!(v & CLIP_8b_MASK)) ? (uint8_t)(v >> DFIX) : (v < 0) ? 0u : 255u; |
77 | } |
78 | #undef CLIP_8b_MASK |
79 | |
80 | // vertical accumulation |
81 | static void VFilter(SmoothParams* const p) { |
82 | const uint8_t* src = p->src_; |
83 | const int w = p->width_; |
84 | uint16_t* const cur = p->cur_; |
85 | const uint16_t* const top = p->top_; |
86 | uint16_t* const out = p->end_; |
87 | uint16_t sum = 0; // all arithmetic is modulo 16bit |
88 | int x; |
89 | |
90 | for (x = 0; x < w; ++x) { |
91 | uint16_t new_value; |
92 | sum += src[x]; |
93 | new_value = top[x] + sum; |
94 | out[x] = new_value - cur[x]; // vertical sum of 'r' pixels. |
95 | cur[x] = new_value; |
96 | } |
97 | // move input pointers one row down |
98 | p->top_ = p->cur_; |
99 | p->cur_ += w; |
100 | if (p->cur_ == p->end_) p->cur_ = p->start_; // roll-over |
101 | // We replicate edges, as it's somewhat easier as a boundary condition. |
102 | // That's why we don't update the 'src' pointer on top/bottom area: |
103 | if (p->row_ >= 0 && p->row_ < p->height_ - 1) { |
104 | p->src_ += p->stride_; |
105 | } |
106 | } |
107 | |
108 | // horizontal accumulation. We use mirror replication of missing pixels, as it's |
109 | // a little easier to implement (surprisingly). |
110 | static void HFilter(SmoothParams* const p) { |
111 | const uint16_t* const in = p->end_; |
112 | uint16_t* const out = p->average_; |
113 | const uint32_t scale = p->scale_; |
114 | const int w = p->width_; |
115 | const int r = p->radius_; |
116 | |
117 | int x; |
118 | for (x = 0; x <= r; ++x) { // left mirroring |
119 | const uint16_t delta = in[x + r - 1] + in[r - x]; |
120 | out[x] = (delta * scale) >> FIX; |
121 | } |
122 | for (; x < w - r; ++x) { // bulk middle run |
123 | const uint16_t delta = in[x + r] - in[x - r - 1]; |
124 | out[x] = (delta * scale) >> FIX; |
125 | } |
126 | for (; x < w; ++x) { // right mirroring |
127 | const uint16_t delta = |
128 | 2 * in[w - 1] - in[2 * w - 2 - r - x] - in[x - r - 1]; |
129 | out[x] = (delta * scale) >> FIX; |
130 | } |
131 | } |
132 | |
133 | // emit one filtered output row |
134 | static void ApplyFilter(SmoothParams* const p) { |
135 | const uint16_t* const average = p->average_; |
136 | const int w = p->width_; |
137 | const int16_t* const correction = p->correction_; |
138 | #if defined(USE_DITHERING) |
139 | const uint8_t* const dither = kOrderedDither[p->row_ % DSIZE]; |
140 | #endif |
141 | uint8_t* const dst = p->dst_; |
142 | int x; |
143 | for (x = 0; x < w; ++x) { |
144 | const int v = dst[x]; |
145 | if (v < p->max_ && v > p->min_) { |
146 | const int c = (v << DFIX) + correction[average[x] - (v << LFIX)]; |
147 | #if defined(USE_DITHERING) |
148 | dst[x] = clip_8b(c + dither[x % DSIZE]); |
149 | #else |
150 | dst[x] = clip_8b(c); |
151 | #endif |
152 | } |
153 | } |
154 | p->dst_ += p->stride_; // advance output pointer |
155 | } |
156 | |
157 | //------------------------------------------------------------------------------ |
158 | // Initialize correction table |
159 | |
160 | static void InitCorrectionLUT(int16_t* const lut, int min_dist) { |
161 | // The correction curve is: |
162 | // f(x) = x for x <= threshold2 |
163 | // f(x) = 0 for x >= threshold1 |
164 | // and a linear interpolation for range x=[threshold2, threshold1] |
165 | // (along with f(-x) = -f(x) symmetry). |
166 | // Note that: threshold2 = 3/4 * threshold1 |
167 | const int threshold1 = min_dist << LFIX; |
168 | const int threshold2 = (3 * threshold1) >> 2; |
169 | const int max_threshold = threshold2 << DFIX; |
170 | const int delta = threshold1 - threshold2; |
171 | int i; |
172 | for (i = 1; i <= LUT_SIZE; ++i) { |
173 | int c = (i <= threshold2) ? (i << DFIX) |
174 | : (i < threshold1) ? max_threshold * (threshold1 - i) / delta |
175 | : 0; |
176 | c >>= LFIX; |
177 | lut[+i] = +c; |
178 | lut[-i] = -c; |
179 | } |
180 | lut[0] = 0; |
181 | } |
182 | |
183 | static void CountLevels(SmoothParams* const p) { |
184 | int i, j, last_level; |
185 | uint8_t used_levels[256] = { 0 }; |
186 | const uint8_t* data = p->src_; |
187 | p->min_ = 255; |
188 | p->max_ = 0; |
189 | for (j = 0; j < p->height_; ++j) { |
190 | for (i = 0; i < p->width_; ++i) { |
191 | const int v = data[i]; |
192 | if (v < p->min_) p->min_ = v; |
193 | if (v > p->max_) p->max_ = v; |
194 | used_levels[v] = 1; |
195 | } |
196 | data += p->stride_; |
197 | } |
198 | // Compute the mininum distance between two non-zero levels. |
199 | p->min_level_dist_ = p->max_ - p->min_; |
200 | last_level = -1; |
201 | for (i = 0; i < 256; ++i) { |
202 | if (used_levels[i]) { |
203 | ++p->num_levels_; |
204 | if (last_level >= 0) { |
205 | const int level_dist = i - last_level; |
206 | if (level_dist < p->min_level_dist_) { |
207 | p->min_level_dist_ = level_dist; |
208 | } |
209 | } |
210 | last_level = i; |
211 | } |
212 | } |
213 | } |
214 | |
215 | // Initialize all params. |
216 | static int InitParams(uint8_t* const data, int width, int height, int stride, |
217 | int radius, SmoothParams* const p) { |
218 | const int R = 2 * radius + 1; // total size of the kernel |
219 | |
220 | const size_t size_scratch_m = (R + 1) * width * sizeof(*p->start_); |
221 | const size_t size_m = width * sizeof(*p->average_); |
222 | const size_t size_lut = (1 + 2 * LUT_SIZE) * sizeof(*p->correction_); |
223 | const size_t total_size = size_scratch_m + size_m + size_lut; |
224 | uint8_t* mem = (uint8_t*)WebPSafeMalloc(1U, total_size); |
225 | |
226 | if (mem == NULL) return 0; |
227 | p->mem_ = (void*)mem; |
228 | |
229 | p->start_ = (uint16_t*)mem; |
230 | p->cur_ = p->start_; |
231 | p->end_ = p->start_ + R * width; |
232 | p->top_ = p->end_ - width; |
233 | memset(p->top_, 0, width * sizeof(*p->top_)); |
234 | mem += size_scratch_m; |
235 | |
236 | p->average_ = (uint16_t*)mem; |
237 | mem += size_m; |
238 | |
239 | p->width_ = width; |
240 | p->height_ = height; |
241 | p->stride_ = stride; |
242 | p->src_ = data; |
243 | p->dst_ = data; |
244 | p->radius_ = radius; |
245 | p->scale_ = (1 << (FIX + LFIX)) / (R * R); // normalization constant |
246 | p->row_ = -radius; |
247 | |
248 | // analyze the input distribution so we can best-fit the threshold |
249 | CountLevels(p); |
250 | |
251 | // correction table |
252 | p->correction_ = ((int16_t*)mem) + LUT_SIZE; |
253 | InitCorrectionLUT(p->correction_, p->min_level_dist_); |
254 | |
255 | return 1; |
256 | } |
257 | |
258 | static void CleanupParams(SmoothParams* const p) { |
259 | WebPSafeFree(p->mem_); |
260 | } |
261 | |
262 | int WebPDequantizeLevels(uint8_t* const data, int width, int height, int stride, |
263 | int strength) { |
264 | int radius = 4 * strength / 100; |
265 | |
266 | if (strength < 0 || strength > 100) return 0; |
267 | if (data == NULL || width <= 0 || height <= 0) return 0; // bad params |
268 | |
269 | // limit the filter size to not exceed the image dimensions |
270 | if (2 * radius + 1 > width) radius = (width - 1) >> 1; |
271 | if (2 * radius + 1 > height) radius = (height - 1) >> 1; |
272 | |
273 | if (radius > 0) { |
274 | SmoothParams p; |
275 | memset(&p, 0, sizeof(p)); |
276 | if (!InitParams(data, width, height, stride, radius, &p)) return 0; |
277 | if (p.num_levels_ > 2) { |
278 | for (; p.row_ < p.height_; ++p.row_) { |
279 | VFilter(&p); // accumulate average of input |
280 | // Need to wait few rows in order to prime the filter, |
281 | // before emitting some output. |
282 | if (p.row_ >= p.radius_) { |
283 | HFilter(&p); |
284 | ApplyFilter(&p); |
285 | } |
286 | } |
287 | } |
288 | CleanupParams(&p); |
289 | } |
290 | return 1; |
291 | } |
292 | |