1 /* NeuQuant Neural-Net Quantization Algorithm
2 * ------------------------------------------
3 *
4 * Copyright (c) 1994 Anthony Dekker
5 *
6 * NEUQUANT Neural-Net quantization algorithm by Anthony Dekker, 1994.
7 * See "Kohonen neural networks for optimal colour quantization"
8 * in "Network: Computation in Neural Systems" Vol. 5 (1994) pp 351-367.
9 * for a discussion of the algorithm.
10 * See also http://members.ozemail.com.au/~dekker/NEUQUANT.HTML
11 *
12 * Any party obtaining a copy of these files from the author, directly or
13 * indirectly, is granted, free of charge, a full and unrestricted irrevocable,
14 * world-wide, paid up, royalty-free, nonexclusive right and license to deal
15 * in this software and documentation files (the "Software"), including without
16 * limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
17 * and/or sell copies of the Software, and to permit persons who receive
18 * copies from any such party to do so, with the only requirement being
19 * that this copyright notice remain intact.
20 *
21 *
22 * Modified to process 32bit RGBA images.
23 * Stuart Coyle 2004-2007
24 * From: http://pngnq.sourceforge.net/
25 *
26 * Ported to libgd by Pierre A. Joye
27 * (and make it thread safety by droping static and global variables)
28 */
29
30 #ifdef HAVE_CONFIG_H
31 #include "config.h"
32 #endif /* HAVE_CONFIG_H */
33
34 #include <stdlib.h>
35 #include <string.h>
36 #include "gd.h"
37 #include "gdhelpers.h"
38 #include "gd_errors.h"
39
40 #include "gd_nnquant.h"
41
42 /* Network Definitions
43 ------------------- */
44
45 #define maxnetpos (MAXNETSIZE-1)
46 #define netbiasshift 4 /* bias for colour values */
47 #define ncycles 100 /* no. of learning cycles */
48
49 /* defs for freq and bias */
50 #define intbiasshift 16 /* bias for fractions */
51 #define intbias (((int) 1)<<intbiasshift)
52 #define gammashift 10 /* gamma = 1024 */
53 #define gamma (((int) 1)<<gammashift)
54 #define betashift 10
55 #define beta (intbias>>betashift) /* beta = 1/1024 */
56 #define betagamma (intbias<<(gammashift-betashift))
57
58 /* defs for decreasing radius factor */
59 #define initrad (MAXNETSIZE>>3) /* for 256 cols, radius starts */
60 #define radiusbiasshift 6 /* at 32.0 biased by 6 bits */
61 #define radiusbias (((int) 1)<<radiusbiasshift)
62 #define initradius (initrad*radiusbias) /* and decreases by a */
63 #define radiusdec 30 /* factor of 1/30 each cycle */
64
65 /* defs for decreasing alpha factor */
66 #define alphabiasshift 10 /* alpha starts at 1.0 */
67 #define initalpha (((int) 1)<<alphabiasshift)
68
69 /* radbias and alpharadbias used for radpower calculation */
70 #define radbiasshift 8
71 #define radbias (((int) 1)<<radbiasshift)
72 #define alpharadbshift (alphabiasshift+radbiasshift)
73 #define alpharadbias (((int) 1)<<alpharadbshift)
74
75 #define ALPHA 0
76 #define RED 1
77 #define BLUE 2
78 #define GREEN 3
79
80 typedef int nq_pixel[5];
81
82 typedef struct {
83 /* biased by 10 bits */
84 int alphadec;
85
86 /* lengthcount = H*W*3 */
87 int lengthcount;
88
89 /* sampling factor 1..30 */
90 int samplefac;
91
92 /* Number of colours to use. Made a global instead of #define */
93 int netsize;
94
95 /* for network lookup - really 256 */
96 int netindex[256];
97
98 /* ABGRc */
99 /* the network itself */
100 nq_pixel network[MAXNETSIZE];
101
102 /* bias and freq arrays for learning */
103 int bias[MAXNETSIZE];
104 int freq[MAXNETSIZE];
105
106 /* radpower for precomputation */
107 int radpower[initrad];
108
109 /* the input image itself */
110 unsigned char *thepicture;
111 } nn_quant;
112
113 /* Initialise network in range (0,0,0,0) to (255,255,255,255) and set parameters
114 ----------------------------------------------------------------------- */
initnet(nnq,thepic,len,sample,colours)115 static void initnet(nnq, thepic, len, sample, colours)
116 nn_quant *nnq;
117 unsigned char *thepic;
118 int len;
119 int sample;
120 int colours;
121 {
122 register int i;
123 register int *p;
124
125 /* Clear out network from previous runs */
126 /* thanks to Chen Bin for this fix */
127 memset((void*)nnq->network, 0, sizeof(nq_pixel)*MAXNETSIZE);
128
129 nnq->thepicture = thepic;
130 nnq->lengthcount = len;
131 nnq->samplefac = sample;
132 nnq->netsize = colours;
133
134 for (i=0; i < nnq->netsize; i++) {
135 p = nnq->network[i];
136 p[0] = p[1] = p[2] = p[3] = (i << (netbiasshift+8)) / nnq->netsize;
137 nnq->freq[i] = intbias / nnq->netsize; /* 1/netsize */
138 nnq->bias[i] = 0;
139 }
140 }
141
142 /* -------------------------- */
143
144 /* Unbias network to give byte values 0..255 and record
145 * position i to prepare for sort
146 */
147 /* -------------------------- */
148
unbiasnet(nn_quant * nnq)149 static void unbiasnet(nn_quant *nnq)
150 {
151 int i,j,temp;
152
153 for (i=0; i < nnq->netsize; i++) {
154 for (j=0; j<4; j++) {
155 /* OLD CODE: network[i][j] >>= netbiasshift; */
156 /* Fix based on bug report by Juergen Weigert jw@suse.de */
157 temp = (nnq->network[i][j] + (1 << (netbiasshift - 1))) >> netbiasshift;
158 if (temp > 255) temp = 255;
159 nnq->network[i][j] = temp;
160 }
161 nnq->network[i][4] = i; /* record colour no */
162 }
163 }
164
165 /* Output colormap to unsigned char ptr in RGBA format */
getcolormap(nnq,map)166 static void getcolormap(nnq, map)
167 nn_quant *nnq;
168 unsigned char *map;
169 {
170 int i,j;
171 for(j=0; j < nnq->netsize; j++) {
172 for (i=3; i>=0; i--) {
173 *map = nnq->network[j][i];
174 map++;
175 }
176 }
177 }
178
179 /* Insertion sort of network and building of netindex[0..255] (to do after unbias)
180 ------------------------------------------------------------------------------- */
inxbuild(nn_quant * nnq)181 static void inxbuild(nn_quant *nnq)
182 {
183 register int i,j,smallpos,smallval;
184 register int *p,*q;
185 int previouscol,startpos;
186
187 previouscol = 0;
188 startpos = 0;
189 for (i=0; i < nnq->netsize; i++) {
190 p = nnq->network[i];
191 smallpos = i;
192 smallval = p[2]; /* index on g */
193 /* find smallest in i..netsize-1 */
194 for (j=i+1; j < nnq->netsize; j++) {
195 q = nnq->network[j];
196 if (q[2] < smallval) { /* index on g */
197 smallpos = j;
198 smallval = q[2]; /* index on g */
199 }
200 }
201 q = nnq->network[smallpos];
202 /* swap p (i) and q (smallpos) entries */
203 if (i != smallpos) {
204 j = q[0];
205 q[0] = p[0];
206 p[0] = j;
207 j = q[1];
208 q[1] = p[1];
209 p[1] = j;
210 j = q[2];
211 q[2] = p[2];
212 p[2] = j;
213 j = q[3];
214 q[3] = p[3];
215 p[3] = j;
216 j = q[4];
217 q[4] = p[4];
218 p[4] = j;
219 }
220 /* smallval entry is now in position i */
221 if (smallval != previouscol) {
222 nnq->netindex[previouscol] = (startpos+i)>>1;
223 for (j=previouscol+1; j<smallval; j++) nnq->netindex[j] = i;
224 previouscol = smallval;
225 startpos = i;
226 }
227 }
228 nnq->netindex[previouscol] = (startpos+maxnetpos)>>1;
229 for (j=previouscol+1; j<256; j++) nnq->netindex[j] = maxnetpos; /* really 256 */
230 }
231
232
233 /* Search for ABGR values 0..255 (after net is unbiased) and return colour index
234 ---------------------------------------------------------------------------- */
inxsearch(nnq,al,b,g,r)235 static unsigned int inxsearch(nnq, al,b,g,r)
236 nn_quant *nnq;
237 register int al, b, g, r;
238 {
239 register int i, j, dist, a, bestd;
240 register int *p;
241 unsigned int best;
242
243 bestd = 1000; /* biggest possible dist is 256*3 */
244 best = 0;
245 i = nnq->netindex[g]; /* index on g */
246 j = i-1; /* start at netindex[g] and work outwards */
247
248 while ((i<nnq->netsize) || (j>=0)) {
249 if (i< nnq->netsize) {
250 p = nnq->network[i];
251 dist = p[2] - g; /* inx key */
252 if (dist >= bestd) i = nnq->netsize; /* stop iter */
253 else {
254 i++;
255 if (dist<0) dist = -dist;
256 a = p[1] - b;
257 if (a<0) a = -a;
258 dist += a;
259 if (dist<bestd) {
260 a = p[3] - r;
261 if (a<0) a = -a;
262 dist += a;
263 }
264 if(dist<bestd) {
265 a = p[0] - al;
266 if (a<0) a = -a;
267 dist += a;
268 }
269 if (dist<bestd) {
270 bestd=dist;
271 best=p[4];
272 }
273 }
274 }
275
276 if (j>=0) {
277 p = nnq->network[j];
278 dist = g - p[2]; /* inx key - reverse dif */
279 if (dist >= bestd) j = -1; /* stop iter */
280 else {
281 j--;
282 if (dist<0) dist = -dist;
283 a = p[1] - b;
284 if (a<0) a = -a;
285 dist += a;
286 if (dist<bestd) {
287 a = p[3] - r;
288 if (a<0) a = -a;
289 dist += a;
290 }
291 if(dist<bestd) {
292 a = p[0] - al;
293 if (a<0) a = -a;
294 dist += a;
295 }
296 if (dist<bestd) {
297 bestd=dist;
298 best=p[4];
299 }
300 }
301 }
302 }
303
304 return(best);
305 }
306
307 /* Search for biased ABGR values
308 ---------------------------- */
contest(nnq,al,b,g,r)309 static int contest(nnq, al,b,g,r)
310 nn_quant *nnq;
311 register int al,b,g,r;
312 {
313 /* finds closest neuron (min dist) and updates freq */
314 /* finds best neuron (min dist-bias) and returns position */
315 /* for frequently chosen neurons, freq[i] is high and bias[i] is negative */
316 /* bias[i] = gamma*((1/netsize)-freq[i]) */
317
318 register int i,dist,a,biasdist,betafreq;
319 unsigned int bestpos,bestbiaspos;
320 double bestd,bestbiasd;
321 register int *p,*f, *n;
322
323 bestd = ~(((int) 1)<<31);
324 bestbiasd = bestd;
325 bestpos = 0;
326 bestbiaspos = bestpos;
327 p = nnq->bias;
328 f = nnq->freq;
329
330 for (i=0; i< nnq->netsize; i++) {
331 n = nnq->network[i];
332 dist = n[0] - al;
333 if (dist<0) dist = -dist;
334 a = n[1] - b;
335 if (a<0) a = -a;
336 dist += a;
337 a = n[2] - g;
338 if (a<0) a = -a;
339 dist += a;
340 a = n[3] - r;
341 if (a<0) a = -a;
342 dist += a;
343 if (dist<bestd) {
344 bestd=dist;
345 bestpos=i;
346 }
347 biasdist = dist - ((*p)>>(intbiasshift-netbiasshift));
348 if (biasdist<bestbiasd) {
349 bestbiasd=biasdist;
350 bestbiaspos=i;
351 }
352 betafreq = (*f >> betashift);
353 *f++ -= betafreq;
354 *p++ += (betafreq<<gammashift);
355 }
356 nnq->freq[bestpos] += beta;
357 nnq->bias[bestpos] -= betagamma;
358 return(bestbiaspos);
359 }
360
361
362 /* Move neuron i towards biased (a,b,g,r) by factor alpha
363 ---------------------------------------------------- */
364
altersingle(nnq,alpha,i,al,b,g,r)365 static void altersingle(nnq, alpha,i,al,b,g,r)
366 nn_quant *nnq;
367 register int alpha,i,al,b,g,r;
368 {
369 register int *n;
370
371 n = nnq->network[i]; /* alter hit neuron */
372 *n -= (alpha*(*n - al)) / initalpha;
373 n++;
374 *n -= (alpha*(*n - b)) / initalpha;
375 n++;
376 *n -= (alpha*(*n - g)) / initalpha;
377 n++;
378 *n -= (alpha*(*n - r)) / initalpha;
379 }
380
381
382 /* Move adjacent neurons by precomputed alpha*(1-((i-j)^2/[r]^2)) in radpower[|i-j|]
383 --------------------------------------------------------------------------------- */
384
alterneigh(nnq,rad,i,al,b,g,r)385 static void alterneigh(nnq, rad,i,al,b,g,r)
386 nn_quant *nnq;
387 int rad,i;
388 register int al,b,g,r;
389 {
390 register int j,k,lo,hi,a;
391 register int *p, *q;
392
393 lo = i-rad;
394 if (lo<-1) lo=-1;
395 hi = i+rad;
396 if (hi>nnq->netsize) hi=nnq->netsize;
397
398 j = i+1;
399 k = i-1;
400 q = nnq->radpower;
401 while ((j<hi) || (k>lo)) {
402 a = (*(++q));
403 if (j<hi) {
404 p = nnq->network[j];
405 *p -= (a*(*p - al)) / alpharadbias;
406 p++;
407 *p -= (a*(*p - b)) / alpharadbias;
408 p++;
409 *p -= (a*(*p - g)) / alpharadbias;
410 p++;
411 *p -= (a*(*p - r)) / alpharadbias;
412 j++;
413 }
414 if (k>lo) {
415 p = nnq->network[k];
416 *p -= (a*(*p - al)) / alpharadbias;
417 p++;
418 *p -= (a*(*p - b)) / alpharadbias;
419 p++;
420 *p -= (a*(*p - g)) / alpharadbias;
421 p++;
422 *p -= (a*(*p - r)) / alpharadbias;
423 k--;
424 }
425 }
426 }
427
428
429 /* Main Learning Loop
430 ------------------ */
431
learn(nnq,verbose)432 static void learn(nnq, verbose) /* Stu: N.B. added parameter so that main() could control verbosity. */
433 nn_quant *nnq;
434 int verbose;
435 {
436 register int i,j,al,b,g,r;
437 int radius,rad,alpha,step,delta,samplepixels;
438 register unsigned char *p;
439 unsigned char *lim;
440
441 nnq->alphadec = 30 + ((nnq->samplefac-1)/3);
442 p = nnq->thepicture;
443 lim = nnq->thepicture + nnq->lengthcount;
444 samplepixels = nnq->lengthcount/(4 * nnq->samplefac);
445 /* here's a problem with small images: samplepixels < ncycles => delta = 0 */
446 delta = samplepixels/ncycles;
447 /* kludge to fix */
448 if(delta==0) delta = 1;
449 alpha = initalpha;
450 radius = initradius;
451
452 rad = radius >> radiusbiasshift;
453
454 for (i=0; i<rad; i++)
455 nnq->radpower[i] = alpha*(((rad*rad - i*i)*radbias)/(rad*rad));
456
457 if (verbose) gd_error_ex(GD_NOTICE, "beginning 1D learning: initial radius=%d\n", rad);
458
459 if ((nnq->lengthcount%prime1) != 0) step = 4*prime1;
460 else {
461 if ((nnq->lengthcount%prime2) !=0) step = 4*prime2;
462 else {
463 if ((nnq->lengthcount%prime3) !=0) step = 4*prime3;
464 else step = 4*prime4;
465 }
466 }
467
468 i = 0;
469 while (i < samplepixels) {
470 al = p[ALPHA] << netbiasshift;
471 b = p[BLUE] << netbiasshift;
472 g = p[GREEN] << netbiasshift;
473 r = p[RED] << netbiasshift;
474 j = contest(nnq, al,b,g,r);
475
476 altersingle(nnq, alpha,j,al,b,g,r);
477 if (rad) alterneigh(nnq, rad,j,al,b,g,r); /* alter neighbours */
478
479 p += step;
480 while (p >= lim) p -= nnq->lengthcount;
481
482 i++;
483 if (i%delta == 0) { /* FPE here if delta=0*/
484 alpha -= alpha / nnq->alphadec;
485 radius -= radius / radiusdec;
486 rad = radius >> radiusbiasshift;
487 if (rad <= 1) rad = 0;
488 for (j=0; j<rad; j++)
489 nnq->radpower[j] = alpha*(((rad*rad - j*j)*radbias)/(rad*rad));
490 }
491 }
492 if (verbose) gd_error_ex(GD_NOTICE, "finished 1D learning: final alpha=%f !\n",((float)alpha)/initalpha);
493 }
494
495 /**
496 * Function: gdImageNeuQuant
497 *
498 * Creates a new palette image from a truecolor image
499 *
500 * This is the same as calling <gdImageCreatePaletteFromTrueColor> with the
501 * quantization method <GD_QUANT_NEUQUANT>.
502 *
503 * Parameters:
504 * im - The image.
505 * max_color - The number of desired palette entries.
506 * sample_factor - The quantization precision between 1 (highest quality) and
507 * 10 (fastest).
508 *
509 * Returns:
510 * A newly create palette image; NULL on failure.
511 */
gdImageNeuQuant(gdImagePtr im,const int max_color,int sample_factor)512 BGD_DECLARE(gdImagePtr) gdImageNeuQuant(gdImagePtr im, const int max_color, int sample_factor)
513 {
514 const int newcolors = max_color;
515 const int verbose = 1;
516
517 int bot_idx, top_idx; /* for remapping of indices */
518 int remap[MAXNETSIZE];
519 int i,x;
520
521 unsigned char map[MAXNETSIZE][4];
522 unsigned char *d;
523
524 nn_quant *nnq = NULL;
525
526 int row;
527 unsigned char *rgba = NULL;
528 gdImagePtr dst = NULL;
529
530 /* Default it to 3 */
531 if (sample_factor < 1) {
532 sample_factor = 3;
533 }
534 /* Start neuquant */
535 /* Pierre:
536 * This implementation works with aligned contiguous buffer only
537 * Upcoming new buffers are contiguous and will be much faster.
538 * let don't bloat this code to support our good "old" 31bit format.
539 * It alos lets us convert palette image, if one likes to reduce
540 * a palette
541 */
542 if (overflow2(gdImageSX(im), gdImageSY(im))
543 || overflow2(gdImageSX(im) * gdImageSY(im), 4)) {
544 goto done;
545 }
546 rgba = (unsigned char *) gdMalloc(gdImageSX(im) * gdImageSY(im) * 4);
547 if (!rgba) {
548 goto done;
549 }
550
551 d = rgba;
552 for (row = 0; row < gdImageSY(im); row++) {
553 int *p = im->tpixels[row];
554 register int c;
555
556 for (i = 0; i < gdImageSX(im); i++) {
557 c = *p;
558 *d++ = gdImageAlpha(im, c);
559 *d++ = gdImageRed(im, c);
560 *d++ = gdImageBlue(im, c);
561 *d++ = gdImageGreen(im, c);
562 p++;
563 }
564 }
565
566 nnq = (nn_quant *) gdMalloc(sizeof(nn_quant));
567 if (!nnq) {
568 goto done;
569 }
570
571 initnet(nnq, rgba, gdImageSY(im) * gdImageSX(im) * 4, sample_factor, newcolors);
572
573 learn(nnq, verbose);
574 unbiasnet(nnq);
575 getcolormap(nnq, (unsigned char*)map);
576 inxbuild(nnq);
577 /* remapping colormap to eliminate opaque tRNS-chunk entries... */
578 for (top_idx = newcolors-1, bot_idx = x = 0; x < newcolors; ++x) {
579 if (map[x][3] == 255) { /* maxval */
580 remap[x] = top_idx--;
581 } else {
582 remap[x] = bot_idx++;
583 }
584 }
585 if (bot_idx != top_idx + 1) {
586 gd_error(" internal logic error: remapped bot_idx = %d, top_idx = %d\n",
587 bot_idx, top_idx);
588 goto done;
589 }
590
591 dst = gdImageCreate(gdImageSX(im), gdImageSY(im));
592 if (!dst) {
593 goto done;
594 }
595
596 for (x = 0; x < newcolors; ++x) {
597 dst->red[remap[x]] = map[x][0];
598 dst->green[remap[x]] = map[x][1];
599 dst->blue[remap[x]] = map[x][2];
600 dst->alpha[remap[x]] = map[x][3];
601 dst->open[remap[x]] = 0;
602 dst->colorsTotal++;
603 }
604
605 /* Do each image row */
606 for ( row = 0; row < gdImageSY(im); ++row ) {
607 int offset;
608 unsigned char *p = dst->pixels[row];
609
610 /* Assign the new colors */
611 offset = row * gdImageSX(im) * 4;
612 for(i=0; i < gdImageSX(im); i++) {
613 p[i] = remap[
614 inxsearch(nnq, rgba[i * 4 + offset + ALPHA],
615 rgba[i * 4 + offset + BLUE],
616 rgba[i * 4 + offset + GREEN],
617 rgba[i * 4 + offset + RED])
618 ];
619 }
620 }
621
622 done:
623 if (rgba) {
624 gdFree(rgba);
625 }
626
627 if (nnq) {
628 gdFree(nnq);
629 }
630 return dst;
631 }
632