1 |
/* Copyright (c) 1994 Regents of the University of California */ |
2 |
|
3 |
#ifndef lint |
4 |
static char SCCSid[] = "$SunId$ LBL"; |
5 |
#endif |
6 |
|
7 |
/* |
8 |
* Neural-Net quantization algorithm based on work of Anthony Dekker |
9 |
*/ |
10 |
|
11 |
#include "standard.h" |
12 |
|
13 |
#include "color.h" |
14 |
|
15 |
#include "random.h" |
16 |
|
17 |
#ifdef COMPAT_MODE |
18 |
#define neu_init new_histo |
19 |
#define neu_pixel cnt_pixel |
20 |
#define neu_colrs cnt_colrs |
21 |
#define neu_clrtab new_clrtab |
22 |
#define neu_map_pixel map_pixel |
23 |
#define neu_map_colrs map_colrs |
24 |
#define neu_dith_colrs dith_colrs |
25 |
#endif |
26 |
/* our color table (global) */ |
27 |
extern BYTE clrtab[256][3]; |
28 |
static int clrtabsiz; |
29 |
|
30 |
#ifndef DEFSMPFAC |
31 |
#ifdef SPEED |
32 |
#define DEFSMPFAC (240/SPEED+3) |
33 |
#else |
34 |
#define DEFSMPFAC 30 |
35 |
#endif |
36 |
#endif |
37 |
|
38 |
int samplefac = DEFSMPFAC; /* sampling factor */ |
39 |
|
40 |
/* Samples array starts off holding spacing between adjacent |
41 |
* samples, and ends up holding actual BGR sample values. |
42 |
*/ |
43 |
static BYTE *thesamples; |
44 |
static int nsamples; |
45 |
static BYTE *cursamp; |
46 |
static long skipcount; |
47 |
|
48 |
#define MAXSKIP (1<<24-1) |
49 |
|
50 |
#define nskip(sp) ((long)(sp)[0]<<16|(long)(sp)[1]<<8|(long)(sp)[2]) |
51 |
|
52 |
#define setskip(sp,n) ((sp)[0]=(n)>>16,(sp)[1]=((n)>>8)&255,(sp)[2]=(n)&255) |
53 |
|
54 |
static cpyclrtab(); |
55 |
|
56 |
|
57 |
neu_init(npixels) /* initialize our sample array */ |
58 |
long npixels; |
59 |
{ |
60 |
register int nsleft; |
61 |
register long sv; |
62 |
double rval, cumprob; |
63 |
long npleft; |
64 |
|
65 |
nsamples = npixels/samplefac; |
66 |
if (nsamples < 600) |
67 |
return(-1); |
68 |
thesamples = (BYTE *)malloc(nsamples*3); |
69 |
if (thesamples == NULL) |
70 |
return(-1); |
71 |
cursamp = thesamples; |
72 |
npleft = npixels; |
73 |
nsleft = nsamples; |
74 |
while (nsleft) { |
75 |
rval = frandom(); /* random distance to next sample */ |
76 |
sv = 0; |
77 |
cumprob = 0.; |
78 |
while ((cumprob += (1.-cumprob)*nsleft/(npleft-sv)) < rval) |
79 |
sv++; |
80 |
if (nsleft == nsamples) |
81 |
skipcount = sv; |
82 |
else { |
83 |
setskip(cursamp, sv); |
84 |
cursamp += 3; |
85 |
} |
86 |
npleft -= sv+1; |
87 |
nsleft--; |
88 |
} |
89 |
setskip(cursamp, npleft); /* tag on end to skip the rest */ |
90 |
cursamp = thesamples; |
91 |
return(0); |
92 |
} |
93 |
|
94 |
|
95 |
neu_pixel(col) /* add pixel to our samples */ |
96 |
register BYTE col[]; |
97 |
{ |
98 |
if (!skipcount--) { |
99 |
skipcount = nskip(cursamp); |
100 |
cursamp[0] = col[BLU]; |
101 |
cursamp[1] = col[GRN]; |
102 |
cursamp[2] = col[RED]; |
103 |
cursamp += 3; |
104 |
} |
105 |
} |
106 |
|
107 |
|
108 |
neu_colrs(cs, n) /* add a scanline to our samples */ |
109 |
register COLR *cs; |
110 |
register int n; |
111 |
{ |
112 |
while (n > skipcount) { |
113 |
cs += skipcount; |
114 |
n -= skipcount+1; |
115 |
skipcount = nskip(cursamp); |
116 |
cursamp[0] = cs[0][BLU]; |
117 |
cursamp[1] = cs[0][GRN]; |
118 |
cursamp[2] = cs[0][RED]; |
119 |
cs++; |
120 |
cursamp += 3; |
121 |
} |
122 |
skipcount -= n; |
123 |
} |
124 |
|
125 |
|
126 |
neu_clrtab(ncolors) /* make new color table using ncolors */ |
127 |
int ncolors; |
128 |
{ |
129 |
clrtabsiz = ncolors; |
130 |
if (clrtabsiz > 256) clrtabsiz = 256; |
131 |
initnet(); |
132 |
learn(); |
133 |
unbiasnet(); |
134 |
cpyclrtab(); |
135 |
inxbuild(); |
136 |
/* we're done with our samples */ |
137 |
free((char *)thesamples); |
138 |
/* reset dithering function */ |
139 |
neu_dith_colrs((BYTE *)NULL, (COLR *)NULL, 0); |
140 |
/* return new color table size */ |
141 |
return(clrtabsiz); |
142 |
} |
143 |
|
144 |
|
145 |
int |
146 |
neu_map_pixel(col) /* get pixel for color */ |
147 |
register BYTE col[]; |
148 |
{ |
149 |
return(inxsearch(col[BLU],col[GRN],col[RED])); |
150 |
} |
151 |
|
152 |
|
153 |
neu_map_colrs(bs, cs, n) /* convert a scanline to color index values */ |
154 |
register BYTE *bs; |
155 |
register COLR *cs; |
156 |
register int n; |
157 |
{ |
158 |
while (n-- > 0) { |
159 |
*bs++ = inxsearch(cs[0][BLU],cs[0][GRN],cs[0][RED]); |
160 |
cs++; |
161 |
} |
162 |
} |
163 |
|
164 |
|
165 |
neu_dith_colrs(bs, cs, n) /* convert scanline to dithered index values */ |
166 |
register BYTE *bs; |
167 |
register COLR *cs; |
168 |
int n; |
169 |
{ |
170 |
static short (*cerr)[3] = NULL; |
171 |
static int N = 0; |
172 |
int err[3], errp[3]; |
173 |
register int x, i; |
174 |
|
175 |
if (n != N) { /* get error propogation array */ |
176 |
if (N) { |
177 |
free((char *)cerr); |
178 |
cerr = NULL; |
179 |
} |
180 |
if (n) |
181 |
cerr = (short (*)[3])malloc(3*n*sizeof(short)); |
182 |
if (cerr == NULL) { |
183 |
N = 0; |
184 |
map_colrs(bs, cs, n); |
185 |
return; |
186 |
} |
187 |
N = n; |
188 |
bzero((char *)cerr, 3*N*sizeof(short)); |
189 |
} |
190 |
err[0] = err[1] = err[2] = 0; |
191 |
for (x = 0; x < n; x++) { |
192 |
for (i = 0; i < 3; i++) { /* dither value */ |
193 |
errp[i] = err[i]; |
194 |
err[i] += cerr[x][i]; |
195 |
#ifdef MAXERR |
196 |
if (err[i] > MAXERR) err[i] = MAXERR; |
197 |
else if (err[i] < -MAXERR) err[i] = -MAXERR; |
198 |
#endif |
199 |
err[i] += cs[x][i]; |
200 |
if (err[i] < 0) err[i] = 0; |
201 |
else if (err[i] > 255) err[i] = 255; |
202 |
} |
203 |
bs[x] = inxsearch(err[BLU],err[GRN],err[RED]); |
204 |
for (i = 0; i < 3; i++) { /* propagate error */ |
205 |
err[i] -= clrtab[bs[x]][i]; |
206 |
err[i] /= 3; |
207 |
cerr[x][i] = err[i] + errp[i]; |
208 |
} |
209 |
} |
210 |
} |
211 |
|
212 |
/* The following was adapted and modified from the original (GW) */ |
213 |
|
214 |
/* cheater definitions (GW) */ |
215 |
#define thepicture thesamples |
216 |
#define lengthcount (nsamples*3) |
217 |
#define samplefac 1 |
218 |
|
219 |
/*----------------------------------------------------------------------*/ |
220 |
/* */ |
221 |
/* NeuQuant */ |
222 |
/* -------- */ |
223 |
/* */ |
224 |
/* Copyright: Anthony Dekker, November 1994 */ |
225 |
/* */ |
226 |
/* This program performs colour quantization of graphics images (SUN */ |
227 |
/* raster files). It uses a Kohonen Neural Network. It produces */ |
228 |
/* better results than existing methods and runs faster, using minimal */ |
229 |
/* space (8kB plus the image itself). The algorithm is described in */ |
230 |
/* the paper "Kohonen Neural Networks for Optimal Colour Quantization" */ |
231 |
/* to appear in the journal "Network: Computation in Neural Systems". */ |
232 |
/* It is a significant improvement of an earlier algorithm. */ |
233 |
/* */ |
234 |
/* This program is distributed free for academic use or for evaluation */ |
235 |
/* by commercial organizations. */ |
236 |
/* */ |
237 |
/* Usage: NeuQuant -n inputfile > outputfile */ |
238 |
/* */ |
239 |
/* where n is a sampling factor for neural learning. */ |
240 |
/* */ |
241 |
/* Program performance compared with other methods is as follows: */ |
242 |
/* */ |
243 |
/* Algorithm | Av. CPU Time | Quantization Error */ |
244 |
/* ------------------------------------------------------------- */ |
245 |
/* NeuQuant -3 | 314 | 5.55 */ |
246 |
/* NeuQuant -10 | 119 | 5.97 */ |
247 |
/* NeuQuant -30 | 65 | 6.53 */ |
248 |
/* Oct-Trees | 141 | 8.96 */ |
249 |
/* Median Cut (XV -best) | 420 | 9.28 */ |
250 |
/* Median Cut (XV -slow) | 72 | 12.15 */ |
251 |
/* */ |
252 |
/* Author's address: Dept of ISCS, National University of Singapore */ |
253 |
/* Kent Ridge, Singapore 0511 */ |
254 |
/* Email: [email protected] */ |
255 |
/*----------------------------------------------------------------------*/ |
256 |
|
257 |
#define bool int |
258 |
#define false 0 |
259 |
#define true 1 |
260 |
|
261 |
/* network defs */ |
262 |
#define netsize clrtabsiz /* number of colours - can change this */ |
263 |
#define maxnetpos (netsize-1) |
264 |
#define netbiasshift 4 /* bias for colour values */ |
265 |
#define ncycles 100 /* no. of learning cycles */ |
266 |
|
267 |
/* defs for freq and bias */ |
268 |
#define intbiasshift 16 /* bias for fractions */ |
269 |
#define intbias (((int) 1)<<intbiasshift) |
270 |
#define gammashift 10 /* gamma = 1024 */ |
271 |
#define gamma (((int) 1)<<gammashift) |
272 |
#define betashift 10 |
273 |
#define beta (intbias>>betashift) /* beta = 1/1024 */ |
274 |
#define betagamma (intbias<<(gammashift-betashift)) |
275 |
|
276 |
/* defs for decreasing radius factor */ |
277 |
#define initrad (256>>3) /* for 256 cols, radius starts */ |
278 |
#define radiusbiasshift 6 /* at 32.0 biased by 6 bits */ |
279 |
#define radiusbias (((int) 1)<<radiusbiasshift) |
280 |
#define initradius (initrad*radiusbias) /* and decreases by a */ |
281 |
#define radiusdec 30 /* factor of 1/30 each cycle */ |
282 |
|
283 |
/* defs for decreasing alpha factor */ |
284 |
#define alphabiasshift 10 /* alpha starts at 1.0 */ |
285 |
#define initalpha (((int) 1)<<alphabiasshift) |
286 |
int alphadec; /* biased by 10 bits */ |
287 |
|
288 |
/* radbias and alpharadbias used for radpower calculation */ |
289 |
#define radbiasshift 8 |
290 |
#define radbias (((int) 1)<<radbiasshift) |
291 |
#define alpharadbshift (alphabiasshift+radbiasshift) |
292 |
#define alpharadbias (((int) 1)<<alpharadbshift) |
293 |
|
294 |
/* four primes near 500 - assume no image has a length so large */ |
295 |
/* that it is divisible by all four primes */ |
296 |
#define prime1 499 |
297 |
#define prime2 491 |
298 |
#define prime3 487 |
299 |
#define prime4 503 |
300 |
|
301 |
typedef int pixel[4]; /* BGRc */ |
302 |
pixel network[256]; |
303 |
|
304 |
int netindex[256]; /* for network lookup - really 256 */ |
305 |
|
306 |
int bias [256]; /* bias and freq arrays for learning */ |
307 |
int freq [256]; |
308 |
int radpower[initrad]; /* radpower for precomputation */ |
309 |
|
310 |
|
311 |
/* initialise network in range (0,0,0) to (255,255,255) */ |
312 |
|
313 |
initnet() |
314 |
{ |
315 |
register int i; |
316 |
register int *p; |
317 |
|
318 |
for (i=0; i<netsize; i++) { |
319 |
p = network[i]; |
320 |
p[0] = p[1] = p[2] = (i << (netbiasshift+8))/netsize; |
321 |
freq[i] = intbias/netsize; /* 1/netsize */ |
322 |
bias[i] = 0; |
323 |
} |
324 |
} |
325 |
|
326 |
|
327 |
/* do after unbias - insertion sort of network and build netindex[0..255] */ |
328 |
|
329 |
inxbuild() |
330 |
{ |
331 |
register int i,j,smallpos,smallval; |
332 |
register int *p,*q; |
333 |
int previouscol,startpos; |
334 |
|
335 |
previouscol = 0; |
336 |
startpos = 0; |
337 |
for (i=0; i<netsize; i++) { |
338 |
p = network[i]; |
339 |
smallpos = i; |
340 |
smallval = p[1]; /* index on g */ |
341 |
/* find smallest in i..netsize-1 */ |
342 |
for (j=i+1; j<netsize; j++) { |
343 |
q = network[j]; |
344 |
if (q[1] < smallval) { /* index on g */ |
345 |
smallpos = j; |
346 |
smallval = q[1]; /* index on g */ |
347 |
} |
348 |
} |
349 |
q = network[smallpos]; |
350 |
/* swap p (i) and q (smallpos) entries */ |
351 |
if (i != smallpos) { |
352 |
j = q[0]; q[0] = p[0]; p[0] = j; |
353 |
j = q[1]; q[1] = p[1]; p[1] = j; |
354 |
j = q[2]; q[2] = p[2]; p[2] = j; |
355 |
j = q[3]; q[3] = p[3]; p[3] = j; |
356 |
} |
357 |
/* smallval entry is now in position i */ |
358 |
if (smallval != previouscol) { |
359 |
netindex[previouscol] = (startpos+i)>>1; |
360 |
for (j=previouscol+1; j<smallval; j++) netindex[j] = i; |
361 |
previouscol = smallval; |
362 |
startpos = i; |
363 |
} |
364 |
} |
365 |
netindex[previouscol] = (startpos+maxnetpos)>>1; |
366 |
for (j=previouscol+1; j<256; j++) netindex[j] = maxnetpos; /* really 256 */ |
367 |
} |
368 |
|
369 |
|
370 |
int inxsearch(b,g,r) /* accepts real BGR values after net is unbiased */ |
371 |
register int b,g,r; |
372 |
{ |
373 |
register int i,j,dist,a,bestd; |
374 |
register int *p; |
375 |
int best; |
376 |
|
377 |
bestd = 1000; /* biggest possible dist is 256*3 */ |
378 |
best = -1; |
379 |
i = netindex[g]; /* index on g */ |
380 |
j = i-1; /* start at netindex[g] and work outwards */ |
381 |
|
382 |
while ((i<netsize) || (j>=0)) { |
383 |
if (i<netsize) { |
384 |
p = network[i]; |
385 |
dist = p[1] - g; /* inx key */ |
386 |
if (dist >= bestd) i = netsize; /* stop iter */ |
387 |
else { |
388 |
i++; |
389 |
if (dist<0) dist = -dist; |
390 |
a = p[0] - b; if (a<0) a = -a; |
391 |
dist += a; |
392 |
if (dist<bestd) { |
393 |
a = p[2] - r; if (a<0) a = -a; |
394 |
dist += a; |
395 |
if (dist<bestd) {bestd=dist; best=p[3];} |
396 |
} |
397 |
} |
398 |
} |
399 |
if (j>=0) { |
400 |
p = network[j]; |
401 |
dist = g - p[1]; /* inx key - reverse dif */ |
402 |
if (dist >= bestd) j = -1; /* stop iter */ |
403 |
else { |
404 |
j--; |
405 |
if (dist<0) dist = -dist; |
406 |
a = p[0] - b; if (a<0) a = -a; |
407 |
dist += a; |
408 |
if (dist<bestd) { |
409 |
a = p[2] - r; if (a<0) a = -a; |
410 |
dist += a; |
411 |
if (dist<bestd) {bestd=dist; best=p[3];} |
412 |
} |
413 |
} |
414 |
} |
415 |
} |
416 |
return(best); |
417 |
} |
418 |
|
419 |
|
420 |
/* finds closest neuron (min dist) and updates freq */ |
421 |
/* finds best neuron (min dist-bias) and returns position */ |
422 |
/* for frequently chosen neurons, freq[i] is high and bias[i] is negative */ |
423 |
/* bias[i] = gamma*((1/netsize)-freq[i]) */ |
424 |
|
425 |
int contest(b,g,r) /* accepts biased BGR values */ |
426 |
register int b,g,r; |
427 |
{ |
428 |
register int i,dist,a,biasdist,betafreq; |
429 |
int bestpos,bestbiaspos,bestd,bestbiasd; |
430 |
register int *p,*f, *n; |
431 |
|
432 |
bestd = ~(((int) 1)<<31); |
433 |
bestbiasd = bestd; |
434 |
bestpos = -1; |
435 |
bestbiaspos = bestpos; |
436 |
p = bias; |
437 |
f = freq; |
438 |
|
439 |
for (i=0; i<netsize; i++) { |
440 |
n = network[i]; |
441 |
dist = n[0] - b; if (dist<0) dist = -dist; |
442 |
a = n[1] - g; if (a<0) a = -a; |
443 |
dist += a; |
444 |
a = n[2] - r; if (a<0) a = -a; |
445 |
dist += a; |
446 |
if (dist<bestd) {bestd=dist; bestpos=i;} |
447 |
biasdist = dist - ((*p)>>(intbiasshift-netbiasshift)); |
448 |
if (biasdist<bestbiasd) {bestbiasd=biasdist; bestbiaspos=i;} |
449 |
betafreq = (*f >> betashift); |
450 |
*f++ -= betafreq; |
451 |
*p++ += (betafreq<<gammashift); |
452 |
} |
453 |
freq[bestpos] += beta; |
454 |
bias[bestpos] -= betagamma; |
455 |
return(bestbiaspos); |
456 |
} |
457 |
|
458 |
|
459 |
/* move neuron i towards (b,g,r) by factor alpha */ |
460 |
|
461 |
altersingle(alpha,i,b,g,r) /* accepts biased BGR values */ |
462 |
register int alpha,i,b,g,r; |
463 |
{ |
464 |
register int *n; |
465 |
|
466 |
n = network[i]; /* alter hit neuron */ |
467 |
*n -= (alpha*(*n - b)) / initalpha; |
468 |
n++; |
469 |
*n -= (alpha*(*n - g)) / initalpha; |
470 |
n++; |
471 |
*n -= (alpha*(*n - r)) / initalpha; |
472 |
} |
473 |
|
474 |
|
475 |
/* move neurons adjacent to i towards (b,g,r) by factor */ |
476 |
/* alpha*(1-((i-j)^2/[r]^2)) precomputed as radpower[|i-j|]*/ |
477 |
|
478 |
alterneigh(rad,i,b,g,r) /* accents biased BGR values */ |
479 |
int rad,i; |
480 |
register int b,g,r; |
481 |
{ |
482 |
register int j,k,lo,hi,a; |
483 |
register int *p, *q; |
484 |
|
485 |
lo = i-rad; if (lo<-1) lo= -1; |
486 |
hi = i+rad; if (hi>netsize) hi=netsize; |
487 |
|
488 |
j = i+1; |
489 |
k = i-1; |
490 |
q = radpower; |
491 |
while ((j<hi) || (k>lo)) { |
492 |
a = (*(++q)); |
493 |
if (j<hi) { |
494 |
p = network[j]; |
495 |
*p -= (a*(*p - b)) / alpharadbias; |
496 |
p++; |
497 |
*p -= (a*(*p - g)) / alpharadbias; |
498 |
p++; |
499 |
*p -= (a*(*p - r)) / alpharadbias; |
500 |
j++; |
501 |
} |
502 |
if (k>lo) { |
503 |
p = network[k]; |
504 |
*p -= (a*(*p - b)) / alpharadbias; |
505 |
p++; |
506 |
*p -= (a*(*p - g)) / alpharadbias; |
507 |
p++; |
508 |
*p -= (a*(*p - r)) / alpharadbias; |
509 |
k--; |
510 |
} |
511 |
} |
512 |
} |
513 |
|
514 |
|
515 |
learn() |
516 |
{ |
517 |
register int i,j,b,g,r; |
518 |
int radius,rad,alpha,step,delta,samplepixels; |
519 |
register unsigned char *p; |
520 |
unsigned char *lim; |
521 |
|
522 |
alphadec = 30 + ((samplefac-1)/3); |
523 |
p = thepicture; |
524 |
lim = thepicture + lengthcount; |
525 |
samplepixels = lengthcount/(3*samplefac); |
526 |
delta = samplepixels/ncycles; |
527 |
alpha = initalpha; |
528 |
radius = initradius; |
529 |
|
530 |
rad = radius >> radiusbiasshift; |
531 |
if (rad <= 1) rad = 0; |
532 |
for (i=0; i<rad; i++) |
533 |
radpower[i] = alpha*(((rad*rad - i*i)*radbias)/(rad*rad)); |
534 |
|
535 |
if ((lengthcount%prime1) != 0) step = 3*prime1; |
536 |
else { |
537 |
if ((lengthcount%prime2) !=0) step = 3*prime2; |
538 |
else { |
539 |
if ((lengthcount%prime3) !=0) step = 3*prime3; |
540 |
else step = 3*prime4; |
541 |
} |
542 |
} |
543 |
|
544 |
i = 0; |
545 |
while (i < samplepixels) { |
546 |
b = p[0] << netbiasshift; |
547 |
g = p[1] << netbiasshift; |
548 |
r = p[2] << netbiasshift; |
549 |
j = contest(b,g,r); |
550 |
|
551 |
altersingle(alpha,j,b,g,r); |
552 |
if (rad) alterneigh(rad,j,b,g,r); /* alter neighbours */ |
553 |
|
554 |
p += step; |
555 |
if (p >= lim) p -= lengthcount; |
556 |
|
557 |
i++; |
558 |
if (i%delta == 0) { |
559 |
alpha -= alpha / alphadec; |
560 |
radius -= radius / radiusdec; |
561 |
rad = radius >> radiusbiasshift; |
562 |
if (rad <= 1) rad = 0; |
563 |
for (j=0; j<rad; j++) |
564 |
radpower[j] = alpha*(((rad*rad - j*j)*radbias)/(rad*rad)); |
565 |
} |
566 |
} |
567 |
} |
568 |
|
569 |
/* unbias network to give 0..255 entries */ |
570 |
/* which can then be used for colour map */ |
571 |
/* and record position i to prepare for sort */ |
572 |
|
573 |
unbiasnet() |
574 |
{ |
575 |
int i,j; |
576 |
|
577 |
for (i=0; i<netsize; i++) { |
578 |
for (j=0; j<3; j++) |
579 |
network[i][j] >>= netbiasshift; |
580 |
network[i][3] = i; /* record colour no */ |
581 |
} |
582 |
} |
583 |
|
584 |
|
585 |
/* Don't do this until the network has been unbiased (GW) */ |
586 |
|
587 |
static |
588 |
cpyclrtab() |
589 |
{ |
590 |
register int i,j,k; |
591 |
|
592 |
for (j=0; j<netsize; j++) { |
593 |
k = network[j][3]; |
594 |
for (i = 0; i < 3; i++) |
595 |
clrtab[k][i] = network[j][2-i]; |
596 |
} |
597 |
} |