• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 package com.bumptech.glide.gifencoder;
2 
3 /*
4  * NeuQuant Neural-Net Quantization Algorithm
5  * ------------------------------------------
6  *
7  * Copyright (c) 1994 Anthony Dekker
8  *
9  * NEUQUANT Neural-Net quantization algorithm by Anthony Dekker, 1994. See
10  * "Kohonen neural networks for optimal colour quantization" in "Network:
11  * Computation in Neural Systems" Vol. 5 (1994) pp 351-367. for a discussion of
12  * the algorithm.
13  *
14  * Any party obtaining a copy of these files from the author, directly or
15  * indirectly, is granted, free of charge, a full and unrestricted irrevocable,
16  * world-wide, paid up, royalty-free, nonexclusive right and license to deal in
17  * this software and documentation files (the "Software"), including without
18  * limitation the rights to use, copy, modify, merge, publish, distribute,
19  * sublicense, and/or sell copies of the Software, and to permit persons who
20  * receive copies from any such party to do so, with the only requirement being
21  * that this copyright notice remain intact.
22  */
23 
24 // Ported to Java 12/00 K Weiner
25 class NeuQuant {
26 
27     protected static final int netsize = 256; /* number of colours used */
28 
29     /* four primes near 500 - assume no image has a length so large */
30   /* that it is divisible by all four primes */
31     protected static final int prime1 = 499;
32 
33     protected static final int prime2 = 491;
34 
35     protected static final int prime3 = 487;
36 
37     protected static final int prime4 = 503;
38 
39     protected static final int minpicturebytes = (3 * prime4);
40 
41   /* minimum size for input image */
42 
43   /*
44    * Program Skeleton ---------------- [select samplefac in range 1..30] [read
45    * image from input file] pic = (unsigned char*) malloc(3*width*height);
46    * initnet(pic,3*width*height,samplefac); learn(); unbiasnet(); [write output
47    * image header, using writecolourmap(f)] inxbuild(); write output image using
48    * inxsearch(b,g,r)
49    */
50 
51   /*
52    * Network Definitions -------------------
53    */
54 
55     protected static final int maxnetpos = (netsize - 1);
56 
57     protected static final int netbiasshift = 4; /* bias for colour values */
58 
59     protected static final int ncycles = 100; /* no. of learning cycles */
60 
61     /* defs for freq and bias */
62     protected static final int intbiasshift = 16; /* bias for fractions */
63 
64     protected static final int intbias = (((int) 1) << intbiasshift);
65 
66     protected static final int gammashift = 10; /* gamma = 1024 */
67 
68     protected static final int gamma = (((int) 1) << gammashift);
69 
70     protected static final int betashift = 10;
71 
72     protected static final int beta = (intbias >> betashift); /* beta = 1/1024 */
73 
74     protected static final int betagamma = (intbias << (gammashift - betashift));
75 
76     /* defs for decreasing radius factor */
77     protected static final int initrad = (netsize >> 3); /*
78                                                          * for 256 cols, radius
79                                                          * starts
80                                                          */
81 
82     protected static final int radiusbiasshift = 6; /* at 32.0 biased by 6 bits */
83 
84     protected static final int radiusbias = (((int) 1) << radiusbiasshift);
85 
86     protected static final int initradius = (initrad * radiusbias); /*
87                                                                    * and
88                                                                    * decreases
89                                                                    * by a
90                                                                    */
91 
92     protected static final int radiusdec = 30; /* factor of 1/30 each cycle */
93 
94     /* defs for decreasing alpha factor */
95     protected static final int alphabiasshift = 10; /* alpha starts at 1.0 */
96 
97     protected static final int initalpha = (((int) 1) << alphabiasshift);
98 
99     protected int alphadec; /* biased by 10 bits */
100 
101     /* radbias and alpharadbias used for radpower calculation */
102     protected static final int radbiasshift = 8;
103 
104     protected static final int radbias = (((int) 1) << radbiasshift);
105 
106     protected static final int alpharadbshift = (alphabiasshift + radbiasshift);
107 
108     protected static final int alpharadbias = (((int) 1) << alpharadbshift);
109 
110   /*
111    * Types and Global Variables --------------------------
112    */
113 
114     protected byte[] thepicture; /* the input image itself */
115 
116     protected int lengthcount; /* lengthcount = H*W*3 */
117 
118     protected int samplefac; /* sampling factor 1..30 */
119 
120     // typedef int pixel[4]; /* BGRc */
121     protected int[][] network; /* the network itself - [netsize][4] */
122 
123     protected int[] netindex = new int[256];
124 
125   /* for network lookup - really 256 */
126 
127     protected int[] bias = new int[netsize];
128 
129     /* bias and freq arrays for learning */
130     protected int[] freq = new int[netsize];
131 
132     protected int[] radpower = new int[initrad];
133 
134   /* radpower for precomputation */
135 
136     /*
137      * Initialise network in range (0,0,0) to (255,255,255) and set parameters
138      * -----------------------------------------------------------------------
139      */
NeuQuant(byte[] thepic, int len, int sample)140     public NeuQuant(byte[] thepic, int len, int sample) {
141 
142         int i;
143         int[] p;
144 
145         thepicture = thepic;
146         lengthcount = len;
147         samplefac = sample;
148 
149         network = new int[netsize][];
150         for (i = 0; i < netsize; i++) {
151             network[i] = new int[4];
152             p = network[i];
153             p[0] = p[1] = p[2] = (i << (netbiasshift + 8)) / netsize;
154             freq[i] = intbias / netsize; /* 1/netsize */
155             bias[i] = 0;
156         }
157     }
158 
colorMap()159     public byte[] colorMap() {
160         byte[] map = new byte[3 * netsize];
161         int[] index = new int[netsize];
162         for (int i = 0; i < netsize; i++)
163             index[network[i][3]] = i;
164         int k = 0;
165         for (int i = 0; i < netsize; i++) {
166             int j = index[i];
167             map[k++] = (byte) (network[j][0]);
168             map[k++] = (byte) (network[j][1]);
169             map[k++] = (byte) (network[j][2]);
170         }
171         return map;
172     }
173 
174     /*
175      * Insertion sort of network and building of netindex[0..255] (to do after
176      * unbias)
177      * -------------------------------------------------------------------------------
178      */
inxbuild()179     public void inxbuild() {
180 
181         int i, j, smallpos, smallval;
182         int[] p;
183         int[] q;
184         int previouscol, startpos;
185 
186         previouscol = 0;
187         startpos = 0;
188         for (i = 0; i < netsize; i++) {
189             p = network[i];
190             smallpos = i;
191             smallval = p[1]; /* index on g */
192       /* find smallest in i..netsize-1 */
193             for (j = i + 1; j < netsize; j++) {
194                 q = network[j];
195                 if (q[1] < smallval) { /* index on g */
196                     smallpos = j;
197                     smallval = q[1]; /* index on g */
198                 }
199             }
200             q = network[smallpos];
201       /* swap p (i) and q (smallpos) entries */
202             if (i != smallpos) {
203                 j = q[0];
204                 q[0] = p[0];
205                 p[0] = j;
206                 j = q[1];
207                 q[1] = p[1];
208                 p[1] = j;
209                 j = q[2];
210                 q[2] = p[2];
211                 p[2] = j;
212                 j = q[3];
213                 q[3] = p[3];
214                 p[3] = j;
215             }
216       /* smallval entry is now in position i */
217             if (smallval != previouscol) {
218                 netindex[previouscol] = (startpos + i) >> 1;
219                 for (j = previouscol + 1; j < smallval; j++)
220                     netindex[j] = i;
221                 previouscol = smallval;
222                 startpos = i;
223             }
224         }
225         netindex[previouscol] = (startpos + maxnetpos) >> 1;
226         for (j = previouscol + 1; j < 256; j++)
227             netindex[j] = maxnetpos; /* really 256 */
228     }
229 
230     /*
231      * Main Learning Loop ------------------
232      */
learn()233     public void learn() {
234 
235         int i, j, b, g, r;
236         int radius, rad, alpha, step, delta, samplepixels;
237         byte[] p;
238         int pix, lim;
239 
240         if (lengthcount < minpicturebytes)
241             samplefac = 1;
242         alphadec = 30 + ((samplefac - 1) / 3);
243         p = thepicture;
244         pix = 0;
245         lim = lengthcount;
246         samplepixels = lengthcount / (3 * samplefac);
247         delta = samplepixels / ncycles;
248         alpha = initalpha;
249         radius = initradius;
250 
251         rad = radius >> radiusbiasshift;
252         if (rad <= 1)
253             rad = 0;
254         for (i = 0; i < rad; i++)
255             radpower[i] = alpha * (((rad * rad - i * i) * radbias) / (rad * rad));
256 
257         // fprintf(stderr,"beginning 1D learning: initial radius=%d\n", rad);
258 
259         if (lengthcount < minpicturebytes)
260             step = 3;
261         else if ((lengthcount % prime1) != 0)
262             step = 3 * prime1;
263         else {
264             if ((lengthcount % prime2) != 0)
265                 step = 3 * prime2;
266             else {
267                 if ((lengthcount % prime3) != 0)
268                     step = 3 * prime3;
269                 else
270                     step = 3 * prime4;
271             }
272         }
273 
274         i = 0;
275         while (i < samplepixels) {
276             b = (p[pix + 0] & 0xff) << netbiasshift;
277             g = (p[pix + 1] & 0xff) << netbiasshift;
278             r = (p[pix + 2] & 0xff) << netbiasshift;
279             j = contest(b, g, r);
280 
281             altersingle(alpha, j, b, g, r);
282             if (rad != 0)
283                 alterneigh(rad, j, b, g, r); /* alter neighbours */
284 
285             pix += step;
286             if (pix >= lim)
287                 pix -= lengthcount;
288 
289             i++;
290             if (delta == 0)
291                 delta = 1;
292             if (i % delta == 0) {
293                 alpha -= alpha / alphadec;
294                 radius -= radius / radiusdec;
295                 rad = radius >> radiusbiasshift;
296                 if (rad <= 1)
297                     rad = 0;
298                 for (j = 0; j < rad; j++)
299                     radpower[j] = alpha * (((rad * rad - j * j) * radbias) / (rad * rad));
300             }
301         }
302         // fprintf(stderr,"finished 1D learning: final alpha=%f
303         // !\n",((float)alpha)/initalpha);
304     }
305 
306     /*
307      * Search for BGR values 0..255 (after net is unbiased) and return colour
308      * index
309      * ----------------------------------------------------------------------------
310      */
map(int b, int g, int r)311     public int map(int b, int g, int r) {
312 
313         int i, j, dist, a, bestd;
314         int[] p;
315         int best;
316 
317         bestd = 1000; /* biggest possible dist is 256*3 */
318         best = -1;
319         i = netindex[g]; /* index on g */
320         j = i - 1; /* start at netindex[g] and work outwards */
321 
322         while ((i < netsize) || (j >= 0)) {
323             if (i < netsize) {
324                 p = network[i];
325                 dist = p[1] - g; /* inx key */
326                 if (dist >= bestd)
327                     i = netsize; /* stop iter */
328                 else {
329                     i++;
330                     if (dist < 0)
331                         dist = -dist;
332                     a = p[0] - b;
333                     if (a < 0)
334                         a = -a;
335                     dist += a;
336                     if (dist < bestd) {
337                         a = p[2] - r;
338                         if (a < 0)
339                             a = -a;
340                         dist += a;
341                         if (dist < bestd) {
342                             bestd = dist;
343                             best = p[3];
344                         }
345                     }
346                 }
347             }
348             if (j >= 0) {
349                 p = network[j];
350                 dist = g - p[1]; /* inx key - reverse dif */
351                 if (dist >= bestd)
352                     j = -1; /* stop iter */
353                 else {
354                     j--;
355                     if (dist < 0)
356                         dist = -dist;
357                     a = p[0] - b;
358                     if (a < 0)
359                         a = -a;
360                     dist += a;
361                     if (dist < bestd) {
362                         a = p[2] - r;
363                         if (a < 0)
364                             a = -a;
365                         dist += a;
366                         if (dist < bestd) {
367                             bestd = dist;
368                             best = p[3];
369                         }
370                     }
371                 }
372             }
373         }
374         return (best);
375     }
376 
process()377     public byte[] process() {
378         learn();
379         unbiasnet();
380         inxbuild();
381         return colorMap();
382     }
383 
384     /*
385      * Unbias network to give byte values 0..255 and record position i to prepare
386      * for sort
387      * -----------------------------------------------------------------------------------
388      */
unbiasnet()389     public void unbiasnet() {
390 
391         int i, j;
392 
393         for (i = 0; i < netsize; i++) {
394             network[i][0] >>= netbiasshift;
395             network[i][1] >>= netbiasshift;
396             network[i][2] >>= netbiasshift;
397             network[i][3] = i; /* record colour no */
398         }
399     }
400 
401     /*
402      * Move adjacent neurons by precomputed alpha*(1-((i-j)^2/[r]^2)) in
403      * radpower[|i-j|]
404      * ---------------------------------------------------------------------------------
405      */
alterneigh(int rad, int i, int b, int g, int r)406     protected void alterneigh(int rad, int i, int b, int g, int r) {
407 
408         int j, k, lo, hi, a, m;
409         int[] p;
410 
411         lo = i - rad;
412         if (lo < -1)
413             lo = -1;
414         hi = i + rad;
415         if (hi > netsize)
416             hi = netsize;
417 
418         j = i + 1;
419         k = i - 1;
420         m = 1;
421         while ((j < hi) || (k > lo)) {
422             a = radpower[m++];
423             if (j < hi) {
424                 p = network[j++];
425                 try {
426                     p[0] -= (a * (p[0] - b)) / alpharadbias;
427                     p[1] -= (a * (p[1] - g)) / alpharadbias;
428                     p[2] -= (a * (p[2] - r)) / alpharadbias;
429                 } catch (Exception e) {
430                 } // prevents 1.3 miscompilation
431             }
432             if (k > lo) {
433                 p = network[k--];
434                 try {
435                     p[0] -= (a * (p[0] - b)) / alpharadbias;
436                     p[1] -= (a * (p[1] - g)) / alpharadbias;
437                     p[2] -= (a * (p[2] - r)) / alpharadbias;
438                 } catch (Exception e) {
439                 }
440             }
441         }
442     }
443 
444     /*
445      * Move neuron i towards biased (b,g,r) by factor alpha
446      * ----------------------------------------------------
447      */
altersingle(int alpha, int i, int b, int g, int r)448     protected void altersingle(int alpha, int i, int b, int g, int r) {
449 
450     /* alter hit neuron */
451         int[] n = network[i];
452         n[0] -= (alpha * (n[0] - b)) / initalpha;
453         n[1] -= (alpha * (n[1] - g)) / initalpha;
454         n[2] -= (alpha * (n[2] - r)) / initalpha;
455     }
456 
457     /*
458      * Search for biased BGR values ----------------------------
459      */
contest(int b, int g, int r)460     protected int contest(int b, int g, int r) {
461 
462     /* finds closest neuron (min dist) and updates freq */
463     /* finds best neuron (min dist-bias) and returns position */
464     /* for frequently chosen neurons, freq[i] is high and bias[i] is negative */
465     /* bias[i] = gamma*((1/netsize)-freq[i]) */
466 
467         int i, dist, a, biasdist, betafreq;
468         int bestpos, bestbiaspos, bestd, bestbiasd;
469         int[] n;
470 
471         bestd = ~(((int) 1) << 31);
472         bestbiasd = bestd;
473         bestpos = -1;
474         bestbiaspos = bestpos;
475 
476         for (i = 0; i < netsize; i++) {
477             n = network[i];
478             dist = n[0] - b;
479             if (dist < 0)
480                 dist = -dist;
481             a = n[1] - g;
482             if (a < 0)
483                 a = -a;
484             dist += a;
485             a = n[2] - r;
486             if (a < 0)
487                 a = -a;
488             dist += a;
489             if (dist < bestd) {
490                 bestd = dist;
491                 bestpos = i;
492             }
493             biasdist = dist - ((bias[i]) >> (intbiasshift - netbiasshift));
494             if (biasdist < bestbiasd) {
495                 bestbiasd = biasdist;
496                 bestbiaspos = i;
497             }
498             betafreq = (freq[i] >> betashift);
499             freq[i] -= betafreq;
500             bias[i] += (betafreq << gammashift);
501         }
502         freq[bestpos] += beta;
503         bias[bestpos] -= betagamma;
504         return (bestbiaspos);
505     }
506 }
507