|
/Users/lyon/j4p/src/ip/gif/neuquantAnimation/NeuQuant.java
|
1 package ip.gif.neuquantAnimation;
2
3 /* NeuQuant Neural-Net Quantization Algorithm
4 * ------------------------------------------
5 *
6 * Copyright (c) 1994 Anthony Dekker
7 *
8 * NEUQUANT Neural-Net quantization algorithm by Anthony Dekker, 1994.
9 * See "Kohonen neural networks for optimal colour quantization"
10 * in "Network: Computation in Neural Systems" Vol. 5 (1994) pp 351-367.
11 * for a discussion of the algorithm.
12 *
13 * Any party obtaining a copy of these files from the author, directly or
14 * indirectly, is granted, free of charge, a full and unrestricted irrevocable,
15 * world-wide, paid up, royalty-free, nonexclusive right and license to deal
16 * in this software and documentation files (the "Software"), including without
17 * limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
18 * and/or sell copies of the Software, and to permit persons who receive
19 * copies from any such party to do so, with the only requirement being
20 * that this copyright notice remain intact.
21 */
22
23 // Ported to Java 12/00 K Weiner
24
25 public 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 protected static final int prime2 = 491;
33 protected static final int prime3 = 487;
34 protected static final int prime4 = 503;
35
36 protected static final int minpicturebytes = (3 * prime4);
37 /* minimum size for input image */
38
39 /* Program Skeleton
40 ----------------
41 [select samplefac in range 1..30]
42 [read image from input file]
43 pic = (unsigned char*) malloc(3*width*height);
44 initnet(pic,3*width*height,samplefac);
45 learn();
46 unbiasnet();
47 [write output image header, using writecolourmap(f)]
48 inxbuild();
49 write output image using inxsearch(b,g,r) */
50
51 /* Network Definitions
52 ------------------- */
53
54 protected static final int maxnetpos = (netsize - 1);
55 protected static final int netbiasshift = 4; /* bias for colour values */
56 protected static final int ncycles = 100; /* no. of learning cycles */
57
58 /* defs for freq and bias */
59 protected static final int intbiasshift = 16; /* bias for fractions */
60 protected static final int intbias = (((int) 1) << intbiasshift);
61 protected static final int gammashift = 10; /* gamma = 1024 */
62 protected static final int gamma = (((int) 1) << gammashift);
63 protected static final int betashift = 10;
64 protected static final int beta = (intbias >> betashift); /* beta = 1/1024 */
65 protected static final int betagamma =
66 (intbias << (gammashift - betashift));
67
68 /* defs for decreasing radius factor */
69 protected static final int initrad = (netsize >> 3); /* for 256 cols, radius starts */
70 protected static final int radiusbiasshift = 6; /* at 32.0 biased by 6 bits */
71 protected static final int radiusbias = (((int) 1) << radiusbiasshift);
72 protected static final int initradius = (initrad * radiusbias); /* and decreases by a */
73 protected static final int radiusdec = 30; /* factor of 1/30 each cycle */
74
75 /* defs for decreasing alpha factor */
76 protected static final int alphabiasshift = 10; /* alpha starts at 1.0 */
77 protected static final int initalpha = (((int) 1) << alphabiasshift);
78
79 protected int alphadec; /* biased by 10 bits */
80
81 /* radbias and alpharadbias used for radpower calculation */
82 protected static final int radbiasshift = 8;
83 protected static final int radbias = (((int) 1) << radbiasshift);
84 protected static final int alpharadbshift = (alphabiasshift + radbiasshift);
85 protected static final int alpharadbias = (((int) 1) << alpharadbshift);
86
87 /* Types and Global Variables
88 -------------------------- */
89
90 protected byte[] thepicture; /* the input image itself */
91 protected int lengthcount; /* lengthcount = H*W*3 */
92
93 protected int samplefac; /* sampling factor 1..30 */
94
95 // typedef int pixel[4]; /* BGRc */
96 protected int[][] network; /* the network itself - [netsize][4] */
97
98 protected int[] netindex = new int[256];
99 /* for network lookup - really 256 */
100
101 protected int[] bias = new int[netsize];
102 /* bias and freq arrays for learning */
103 protected int[] freq = new int[netsize];
104 protected int[] radpower = new int[initrad];
105 /* radpower for precomputation */
106
107 /* Initialise network in range (0,0,0) to (255,255,255) and set parameters
108 ----------------------------------------------------------------------- */
109 public NeuQuant(byte[] thepic, int len, int sample) {
110
111 int i;
112 int[] p;
113
114 thepicture = thepic;
115 lengthcount = len;
116 samplefac = sample;
117
118 network = new int[netsize][];
119 for (i = 0; i < netsize; i++) {
120 network[i] = new int[4];
121 p = network[i];
122 p[0] = p[1] = p[2] = (i << (netbiasshift + 8)) / netsize;
123 freq[i] = intbias / netsize; /* 1/netsize */
124 bias[i] = 0;
125 }
126 }
127
128 public byte[] colorMap() {
129 byte[] map = new byte[3 * netsize];
130 int[] index = new int[netsize];
131 for (int i = 0; i < netsize; i++)
132 index[network[i][3]] = i;
133 int k = 0;
134 for (int i = 0; i < netsize; i++) {
135 int j = index[i];
136 map[k++] = (byte) (network[j][0]);
137 map[k++] = (byte) (network[j][1]);
138 map[k++] = (byte) (network[j][2]);
139 }
140 return map;
141 }
142
143 /* Insertion sort of network and building of netindex[0..255] (to do after unbias)
144 ------------------------------------------------------------------------------- */
145 public void inxbuild() {
146
147 int i, j, smallpos, smallval;
148 int[] p;
149 int[] q;
150 int previouscol, startpos;
151
152 previouscol = 0;
153 startpos = 0;
154 for (i = 0; i < netsize; i++) {
155 p = network[i];
156 smallpos = i;
157 smallval = p[1]; /* index on g */
158 /* find smallest in i..netsize-1 */
159 for (j = i + 1; j < netsize; j++) {
160 q = network[j];
161 if (q[1] < smallval) { /* index on g */
162 smallpos = j;
163 smallval = q[1]; /* index on g */
164 }
165 }
166 q = network[smallpos];
167 /* swap p (i) and q (smallpos) entries */
168 if (i != smallpos) {
169 j = q[0];
170 q[0] = p[0];
171 p[0] = j;
172 j = q[1];
173 q[1] = p[1];
174 p[1] = j;
175 j = q[2];
176 q[2] = p[2];
177 p[2] = j;
178 j = q[3];
179 q[3] = p[3];
180 p[3] = j;
181 }
182 /* smallval entry is now in position i */
183 if (smallval != previouscol) {
184 netindex[previouscol] = (startpos + i) >> 1;
185 for (j = previouscol + 1; j < smallval; j++)
186 netindex[j] = i;
187 previouscol = smallval;
188 startpos = i;
189 }
190 }
191 netindex[previouscol] = (startpos + maxnetpos) >> 1;
192 for (j = previouscol + 1; j < 256; j++)
193 netindex[j] = maxnetpos; /* really 256 */
194 }
195
196 /* Main Learning Loop
197 ------------------ */
198 public void learn() {
199
200 int i, j, b, g, r;
201 int radius, rad, alpha, step, delta, samplepixels;
202 byte[] p;
203 int pix, lim;
204
205 if (lengthcount < minpicturebytes)
206 samplefac = 1;
207 alphadec = 30 + ((samplefac - 1) / 3);
208 p = thepicture;
209 pix = 0;
210 lim = lengthcount;
211 samplepixels = lengthcount / (3 * samplefac);
212 delta = samplepixels / ncycles;
213 alpha = initalpha;
214 radius = initradius;
215
216 rad = radius >> radiusbiasshift;
217 if (rad <= 1)
218 rad = 0;
219 for (i = 0; i < rad; i++)
220 radpower[i] =
221 alpha * (((rad * rad - i * i) * radbias) / (rad * rad));
222
223 //fprintf(stderr,"beginning 1D learning: initial radius=%d\n", rad);
224
225 if (lengthcount < minpicturebytes)
226 step = 3;
227 else if ((lengthcount % prime1) != 0)
228 step = 3 * prime1;
229 else {
230 if ((lengthcount % prime2) != 0)
231 step = 3 * prime2;
232 else {
233 if ((lengthcount % prime3) != 0)
234 step = 3 * prime3;
235 else
236 step = 3 * prime4;
237 }
238 }
239
240 i = 0;
241 while (i < samplepixels) {
242 b = (p[pix + 0] & 0xff) << netbiasshift;
243 g = (p[pix + 1] & 0xff) << netbiasshift;
244 r = (p[pix + 2] & 0xff) << netbiasshift;
245 j = contest(b, g, r);
246
247 altersingle(alpha, j, b, g, r);
248 if (rad != 0)
249 alterneigh(rad, j, b, g, r); /* alter neighbours */
250
251 pix += step;
252 if (pix >= lim)
253 pix -= lengthcount;
254
255 i++;
256 if (delta == 0)
257 delta = 1;
258 if (i % delta == 0) {
259 alpha -= alpha / alphadec;
260 radius -= radius / radiusdec;
261 rad = radius >> radiusbiasshift;
262 if (rad <= 1)
263 rad = 0;
264 for (j = 0; j < rad; j++)
265 radpower[j] =
266 alpha * (((rad * rad - j * j) * radbias) / (rad * rad));
267 }
268 }
269 //fprintf(stderr,"finished 1D learning: final alpha=%f !\n",((float)alpha)/initalpha);
270 }
271
272 /* Search for BGR values 0..255 (after net is unbiased) and return colour index
273 ---------------------------------------------------------------------------- */
274 public int map(int b, int g, int r) {
275
276 int i, j, dist, a, bestd;
277 int[] p;
278 int best;
279
280 bestd = 1000; /* biggest possible dist is 256*3 */
281 best = -1;
282 i = netindex[g]; /* index on g */
283 j = i - 1; /* start at netindex[g] and work outwards */
284
285 while ((i < netsize) || (j >= 0)) {
286 if (i < netsize) {
287 p = network[i];
288 dist = p[1] - g; /* inx key */
289 if (dist >= bestd)
290 i = netsize; /* stop iter */
291 else {
292 i++;
293 if (dist < 0)
294 dist = -dist;
295 a = p[0] - b;
296 if (a < 0)
297 a = -a;
298 dist += a;
299 if (dist < bestd) {
300 a = p[2] - r;
301 if (a < 0)
302 a = -a;
303 dist += a;
304 if (dist < bestd) {
305 bestd = dist;
306 best = p[3];
307 }
308 }
309 }
310 }
311 if (j >= 0) {
312 p = network[j];
313 dist = g - p[1]; /* inx key - reverse dif */
314 if (dist >= bestd)
315 j = -1; /* stop iter */
316 else {
317 j--;
318 if (dist < 0)
319 dist = -dist;
320 a = p[0] - b;
321 if (a < 0)
322 a = -a;
323 dist += a;
324 if (dist < bestd) {
325 a = p[2] - r;
326 if (a < 0)
327 a = -a;
328 dist += a;
329 if (dist < bestd) {
330 bestd = dist;
331 best = p[3];
332 }
333 }
334 }
335 }
336 }
337 return (best);
338 }
339 public byte[] process() {
340 learn();
341 unbiasnet();
342 inxbuild();
343 return colorMap();
344 }
345
346 /* Unbias network to give byte values 0..255 and record position i to prepare for sort
347 ----------------------------------------------------------------------------------- */
348 public void unbiasnet() {
349
350 int i, j;
351
352 for (i = 0; i < netsize; i++) {
353 network[i][0] >>= netbiasshift;
354 network[i][1] >>= netbiasshift;
355 network[i][2] >>= netbiasshift;
356 network[i][3] = i; /* record colour no */
357 }
358 }
359
360 /* Move adjacent neurons by precomputed alpha*(1-((i-j)^2/[r]^2)) in radpower[|i-j|]
361 --------------------------------------------------------------------------------- */
362 protected void alterneigh(int rad, int i, int b, int g, int r) {
363
364 int j, k, lo, hi, a, m;
365 int[] p;
366
367 lo = i - rad;
368 if (lo < -1)
369 lo = -1;
370 hi = i + rad;
371 if (hi > netsize)
372 hi = netsize;
373
374 j = i + 1;
375 k = i - 1;
376 m = 1;
377 while ((j < hi) || (k > lo)) {
378 a = radpower[m++];
379 if (j < hi) {
380 p = network[j++];
381 try {
382 p[0] -= (a * (p[0] - b)) / alpharadbias;
383 p[1] -= (a * (p[1] - g)) / alpharadbias;
384 p[2] -= (a * (p[2] - r)) / alpharadbias;
385 } catch (Exception e) {
386 } // prevents 1.3 miscompilation
387 }
388 if (k > lo) {
389 p = network[k--];
390 try {
391 p[0] -= (a * (p[0] - b)) / alpharadbias;
392 p[1] -= (a * (p[1] - g)) / alpharadbias;
393 p[2] -= (a * (p[2] - r)) / alpharadbias;
394 } catch (Exception e) {
395 }
396 }
397 }
398 }
399
400 /* Move neuron i towards biased (b,g,r) by factor alpha
401 ---------------------------------------------------- */
402 protected void altersingle(int alpha, int i, int b, int g, int r) {
403
404 /* alter hit neuron */
405 int[] n = network[i];
406 n[0] -= (alpha * (n[0] - b)) / initalpha;
407 n[1] -= (alpha * (n[1] - g)) / initalpha;
408 n[2] -= (alpha * (n[2] - r)) / initalpha;
409 }
410
411 /* Search for biased BGR values
412 ---------------------------- */
413 protected int contest(int b, int g, int r) {
414
415 /* finds closest neuron (min dist) and updates freq */
416 /* finds best neuron (min dist-bias) and returns position */
417 /* for frequently chosen neurons, freq[i] is high and bias[i] is negative */
418 /* bias[i] = gamma*((1/netsize)-freq[i]) */
419
420 int i, dist, a, biasdist, betafreq;
421 int bestpos, bestbiaspos, bestd, bestbiasd;
422 int[] n;
423
424 bestd = ~(((int) 1) << 31);
425 bestbiasd = bestd;
426 bestpos = -1;
427 bestbiaspos = bestpos;
428
429 for (i = 0; i < netsize; i++) {
430 n = network[i];
431 dist = n[0] - b;
432 if (dist < 0)
433 dist = -dist;
434 a = n[1] - g;
435 if (a < 0)
436 a = -a;
437 dist += a;
438 a = n[2] - r;
439 if (a < 0)
440 a = -a;
441 dist += a;
442 if (dist < bestd) {
443 bestd = dist;
444 bestpos = i;
445 }
446 biasdist = dist - ((bias[i]) >> (intbiasshift - netbiasshift));
447 if (biasdist < bestbiasd) {
448 bestbiasd = biasdist;
449 bestbiaspos = i;
450 }
451 betafreq = (freq[i] >> betashift);
452 freq[i] -= betafreq;
453 bias[i] += (betafreq << gammashift);
454 }
455 freq[bestpos] += beta;
456 bias[bestpos] -= betagamma;
457 return (bestbiaspos);
458 }
459 }
460