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 |
|
55 |
neu_init(npixels) /* initialize our sample array */ |
56 |
long npixels; |
57 |
{ |
58 |
register int nsleft; |
59 |
register long sv; |
60 |
double rval, cumprob; |
61 |
long npleft; |
62 |
|
63 |
nsamples = npixels/samplefac; |
64 |
if (nsamples < 600) |
65 |
return(-1); |
66 |
thesamples = (BYTE *)malloc((nsamples+1)*3); |
67 |
if (thesamples == NULL) |
68 |
return(-1); |
69 |
cursamp = thesamples; |
70 |
npleft = npixels; |
71 |
nsleft = nsamples; |
72 |
while (nsleft) { |
73 |
rval = frandom(); /* random distance to next sample */ |
74 |
sv = 0; |
75 |
cumprob = 0.; |
76 |
while ((cumprob += (1.-cumprob)*nsleft/(npleft-sv)) < rval) |
77 |
sv++; |
78 |
setskip(cursamp, sv); |
79 |
cursamp += 3; |
80 |
npleft -= sv; |
81 |
nsleft--; |
82 |
} |
83 |
setskip(cursamp, 0); /* dummy tagged onto end */ |
84 |
cursamp = thesamples; |
85 |
skipcount = nskip(cursamp); |
86 |
return(0); |
87 |
} |
88 |
|
89 |
|
90 |
neu_pixel(col) /* add pixel to our samples */ |
91 |
register BYTE col[]; |
92 |
{ |
93 |
if (!skipcount--) { |
94 |
cursamp[0] = col[BLU]; |
95 |
cursamp[1] = col[GRN]; |
96 |
cursamp[2] = col[RED]; |
97 |
cursamp += 3; |
98 |
skipcount = nskip(cursamp); |
99 |
} |
100 |
} |
101 |
|
102 |
|
103 |
neu_colrs(cs, n) /* add a scanline to our samples */ |
104 |
register COLR *cs; |
105 |
register int n; |
106 |
{ |
107 |
while (n > skipcount) { |
108 |
cs += skipcount; |
109 |
cursamp[0] = cs[0][BLU]; |
110 |
cursamp[1] = cs[0][GRN]; |
111 |
cursamp[2] = cs[0][RED]; |
112 |
cs++; |
113 |
n -= skipcount+1; |
114 |
cursamp += 3; |
115 |
skipcount = nskip(cursamp); |
116 |
} |
117 |
skipcount -= n; |
118 |
} |
119 |
|
120 |
|
121 |
neu_clrtab(ncolors) /* make new color table using ncolors */ |
122 |
int ncolors; |
123 |
{ |
124 |
clrtabsiz = ncolors; |
125 |
if (clrtabsiz > 256) clrtabsiz = 256; |
126 |
initnet(); |
127 |
learn(); |
128 |
unbiasnet(); |
129 |
cpyclrtab(); |
130 |
inxbuild(); |
131 |
/* we're done with our samples */ |
132 |
free((char *)thesamples); |
133 |
/* reset dithering function */ |
134 |
neu_dith_colrs((BYTE *)NULL, (COLR *)NULL, 0); |
135 |
/* return new color table size */ |
136 |
return(clrtabsiz); |
137 |
} |
138 |
|
139 |
|
140 |
int |
141 |
neu_map_pixel(col) /* get pixel for color */ |
142 |
register BYTE col[]; |
143 |
{ |
144 |
return(inxsearch(col[BLU],col[GRN],col[RED])); |
145 |
} |
146 |
|
147 |
|
148 |
neu_map_colrs(bs, cs, n) /* convert a scanline to color index values */ |
149 |
register BYTE *bs; |
150 |
register COLR *cs; |
151 |
register int n; |
152 |
{ |
153 |
while (n-- > 0) { |
154 |
*bs++ = inxsearch(cs[0][BLU],cs[0][GRN],cs[0][RED]); |
155 |
cs++; |
156 |
} |
157 |
} |
158 |
|
159 |
|
160 |
neu_dith_colrs(bs, cs, n) /* convert scanline to dithered index values */ |
161 |
register BYTE *bs; |
162 |
register COLR *cs; |
163 |
int n; |
164 |
{ |
165 |
static short (*cerr)[3] = NULL; |
166 |
static int N = 0; |
167 |
int err[3], errp[3]; |
168 |
register int x, i; |
169 |
|
170 |
if (n != N) { /* get error propogation array */ |
171 |
if (N) { |
172 |
free((char *)cerr); |
173 |
cerr = NULL; |
174 |
} |
175 |
if (n) |
176 |
cerr = (short (*)[3])malloc(3*n*sizeof(short)); |
177 |
if (cerr == NULL) { |
178 |
N = 0; |
179 |
map_colrs(bs, cs, n); |
180 |
return; |
181 |
} |
182 |
N = n; |
183 |
bzero((char *)cerr, 3*N*sizeof(short)); |
184 |
} |
185 |
err[0] = err[1] = err[2] = 0; |
186 |
for (x = 0; x < n; x++) { |
187 |
for (i = 0; i < 3; i++) { /* dither value */ |
188 |
errp[i] = err[i]; |
189 |
err[i] += cerr[x][i]; |
190 |
#ifdef MAXERR |
191 |
if (err[i] > MAXERR) err[i] = MAXERR; |
192 |
else if (err[i] < -MAXERR) err[i] = -MAXERR; |
193 |
#endif |
194 |
err[i] += cs[x][i]; |
195 |
if (err[i] < 0) err[i] = 0; |
196 |
else if (err[i] > 255) err[i] = 255; |
197 |
} |
198 |
bs[x] = inxsearch(err[BLU],err[GRN],err[RED]); |
199 |
for (i = 0; i < 3; i++) { /* propagate error */ |
200 |
err[i] -= clrtab[bs[x]][i]; |
201 |
err[i] /= 3; |
202 |
cerr[x][i] = err[i] + errp[i]; |
203 |
} |
204 |
} |
205 |
} |
206 |
|
207 |
/* The following was adapted and modified from the original (GW) */ |
208 |
/*----------------------------------------------------------------------*/ |
209 |
/* */ |
210 |
/* NeuQuant */ |
211 |
/* -------- */ |
212 |
/* */ |
213 |
/* Copyright: Anthony Dekker, June 1994 */ |
214 |
/* */ |
215 |
/* This program performs colour quantization of graphics images (SUN */ |
216 |
/* raster files). It uses a Kohonen Neural Network. It produces */ |
217 |
/* better results than existing methods and runs faster, using minimal */ |
218 |
/* space (8kB plus the image itself). The algorithm is described in */ |
219 |
/* the paper "Kohonen Neural Networks for Optimal Colour Quantization" */ |
220 |
/* to appear in the journal "Network: Computation in Neural Systems". */ |
221 |
/* It is a significant improvement of an earlier algorithm. */ |
222 |
/* */ |
223 |
/* This program is distributed free for academic use or for evaluation */ |
224 |
/* by commercial organizations. */ |
225 |
/* */ |
226 |
/* Usage: NeuQuant -n inputfile > outputfile */ |
227 |
/* */ |
228 |
/* where n is a sampling factor for neural learning. */ |
229 |
/* */ |
230 |
/* Program performance compared with other methods is as follows: */ |
231 |
/* */ |
232 |
/* Algorithm | Av. CPU Time | Quantization Error */ |
233 |
/* ------------------------------------------------------------- */ |
234 |
/* NeuQuant -3 | 314 | 5.55 */ |
235 |
/* NeuQuant -10 | 119 | 5.97 */ |
236 |
/* NeuQuant -30 | 65 | 6.53 */ |
237 |
/* Oct-Trees | 141 | 8.96 */ |
238 |
/* Median Cut (XV -best) | 420 | 9.28 */ |
239 |
/* Median Cut (XV -slow) | 72 | 12.15 */ |
240 |
/* */ |
241 |
/* Author's address: Dept of ISCS, National University of Singapore */ |
242 |
/* Kent Ridge, Singapore 0511 */ |
243 |
/* Email: [email protected] */ |
244 |
/*----------------------------------------------------------------------*/ |
245 |
|
246 |
#define bool int |
247 |
#define false 0 |
248 |
#define true 1 |
249 |
|
250 |
#define initrad 32 |
251 |
#define radiusdec 30 |
252 |
#define alphadec; 30 |
253 |
|
254 |
/* defs for freq and bias */ |
255 |
#define gammashift 10 |
256 |
#define betashift gammashift |
257 |
#define intbiasshift 16 |
258 |
#define intbias (1<<intbiasshift) |
259 |
#define gamma (1<<gammashift) |
260 |
#define beta (intbias>>betashift) |
261 |
#define betagamma (intbias<<(gammashift-betashift)) |
262 |
#define gammaphi (intbias<<(gammashift-8)) |
263 |
|
264 |
/* defs for rad and alpha */ |
265 |
#define maxrad (initrad+1) |
266 |
#define radiusbiasshift 6 |
267 |
#define radiusbias (1<<radiusbiasshift) |
268 |
#define initradius ((int) (initrad*radiusbias)) |
269 |
#define alphabiasshift 10 |
270 |
#define initalpha (1<<alphabiasshift) |
271 |
#define radbiasshift 8 |
272 |
#define radbias (1<<radbiasshift) |
273 |
#define alpharadbshift (alphabiasshift+radbiasshift) |
274 |
#define alpharadbias (1<<alpharadbshift) |
275 |
|
276 |
/* other defs */ |
277 |
#define netbiasshift 4 |
278 |
#define funnyshift (intbiasshift-netbiasshift) |
279 |
#define maxnetval ((256<<netbiasshift)-1) |
280 |
#define ncycles 100 |
281 |
#define jump1 499 /* prime */ |
282 |
#define jump2 491 /* prime */ |
283 |
#define jump3 487 /* any pic whose size was divisible by all */ |
284 |
#define jump4 503 /* four primes would be simply enormous */ |
285 |
|
286 |
/* cheater definitions (GW) */ |
287 |
#define thepicture thesamples |
288 |
#define lengthcount (nsamples*3) |
289 |
#define samplefac 1 |
290 |
|
291 |
typedef int pixel[4]; /* BGRc */ |
292 |
|
293 |
static pixel network[256]; |
294 |
|
295 |
static int netindex[256]; |
296 |
|
297 |
static int bias [256]; |
298 |
static int freq [256]; |
299 |
static int radpower[256]; /* actually need only go up to maxrad */ |
300 |
|
301 |
/* fixed space overhead 256*4+256+256+256+256 words = 256*8 = 8kB */ |
302 |
|
303 |
|
304 |
static |
305 |
initnet() |
306 |
{ |
307 |
register int i; |
308 |
register int *p; |
309 |
|
310 |
for (i=0; i<clrtabsiz; i++) { |
311 |
p = network[i]; |
312 |
p[0] = i << netbiasshift; |
313 |
p[1] = i << netbiasshift; |
314 |
p[2] = i << netbiasshift; |
315 |
freq[i] = intbias >> 8; /* 1/256 */ |
316 |
bias[i] = 0; |
317 |
} |
318 |
} |
319 |
|
320 |
|
321 |
static |
322 |
inxbuild() |
323 |
{ |
324 |
register int i,j,smallpos,smallval; |
325 |
register int *p,*q; |
326 |
int start,previous; |
327 |
|
328 |
previous = 0; |
329 |
start = 0; |
330 |
for (i=0; i<clrtabsiz; i++) { |
331 |
p = network[i]; |
332 |
smallpos = i; |
333 |
smallval = p[1]; /* index on g */ |
334 |
/* find smallest in i+1..clrtabsiz-1 */ |
335 |
for (j=i+1; j<clrtabsiz; j++) { |
336 |
q = network[j]; |
337 |
if (q[1] < smallval) { /* index on g */ |
338 |
smallpos = j; |
339 |
smallval = q[1]; /* index on g */ |
340 |
} |
341 |
} |
342 |
q = network[smallpos]; |
343 |
if (i != smallpos) { |
344 |
j = q[0]; q[0] = p[0]; p[0] = j; |
345 |
j = q[1]; q[1] = p[1]; p[1] = j; |
346 |
j = q[2]; q[2] = p[2]; p[2] = j; |
347 |
j = q[3]; q[3] = p[3]; p[3] = j; |
348 |
} |
349 |
/* smallval entry is now in position i */ |
350 |
if (smallval != previous) { |
351 |
netindex[previous] = (start+i)>>1; |
352 |
for (j=previous+1; j<smallval; j++) netindex[j] = i; |
353 |
previous = smallval; |
354 |
start = i; |
355 |
} |
356 |
} |
357 |
netindex[previous] = (start+clrtabsiz-1)>>1; |
358 |
for (j=previous+1; j<clrtabsiz; j++) netindex[j] = clrtabsiz-1; |
359 |
} |
360 |
|
361 |
|
362 |
static int |
363 |
inxsearch(b,g,r) /* accepts real BGR values after net is unbiased */ |
364 |
register int b,g,r; |
365 |
{ |
366 |
register int i,j,best,x,y,bestd; |
367 |
register int *p; |
368 |
|
369 |
bestd = 1000; /* biggest possible dist is 256*3 */ |
370 |
best = -1; |
371 |
i = netindex[g]; /* index on g */ |
372 |
j = i-1; |
373 |
|
374 |
while ((i<clrtabsiz) || (j>=0)) { |
375 |
if (i<clrtabsiz) { |
376 |
p = network[i]; |
377 |
x = p[1] - g; /* inx key */ |
378 |
if (x >= bestd) i = clrtabsiz; /* stop iter */ |
379 |
else { |
380 |
i++; |
381 |
if (x<0) x = -x; |
382 |
y = p[0] - b; |
383 |
if (y<0) y = -y; |
384 |
x += y; |
385 |
if (x<bestd) { |
386 |
y = p[2] - r; |
387 |
if (y<0) y = -y; |
388 |
x += y; /* x holds distance */ |
389 |
if (x<bestd) {bestd=x; best=p[3];} |
390 |
} |
391 |
} |
392 |
} |
393 |
if (j>=0) { |
394 |
p = network[j]; |
395 |
x = g - p[1]; /* inx key - reverse dif */ |
396 |
if (x >= bestd) j = -1; /* stop iter */ |
397 |
else { |
398 |
j--; |
399 |
if (x<0) x = -x; |
400 |
y = p[0] - b; |
401 |
if (y<0) y = -y; |
402 |
x += y; |
403 |
if (x<bestd) { |
404 |
y = p[2] - r; |
405 |
if (y<0) y = -y; |
406 |
x += y; /* x holds distance */ |
407 |
if (x<bestd) {bestd=x; best=p[3];} |
408 |
} |
409 |
} |
410 |
} |
411 |
} |
412 |
return(best); |
413 |
} |
414 |
|
415 |
|
416 |
static int |
417 |
contest(b,g,r) /* accepts biased BGR values */ |
418 |
register int b,g,r; |
419 |
{ |
420 |
register int i,best,bestbias,x,y,bestd,bestbiasd; |
421 |
register int *p,*q, *pp; |
422 |
|
423 |
bestd = ~(1<<31); |
424 |
bestbiasd = bestd; |
425 |
best = -1; |
426 |
bestbias = best; |
427 |
q = bias; |
428 |
p = freq; |
429 |
for (i=0; i<clrtabsiz; i++) { |
430 |
pp = network[i]; |
431 |
x = pp[0] - b; |
432 |
if (x<0) x = -x; |
433 |
y = pp[1] - g; |
434 |
if (y<0) y = -y; |
435 |
x += y; |
436 |
y = pp[2] - r; |
437 |
if (y<0) y = -y; |
438 |
x += y; /* x holds distance */ |
439 |
/* >> netbiasshift not needed if funnyshift used */ |
440 |
if (x<bestd) {bestd=x; best=i;} |
441 |
y = x - ((*q)>>funnyshift); /* y holds biasd */ |
442 |
if (y<bestbiasd) {bestbiasd=y; bestbias=i;} |
443 |
y = (*p >> betashift); /* y holds beta*freq */ |
444 |
*p -= y; |
445 |
*q += (y<<gammashift); |
446 |
p++; |
447 |
q++; |
448 |
} |
449 |
freq[best] += beta; |
450 |
bias[best] -= betagamma; |
451 |
return(bestbias); |
452 |
} |
453 |
|
454 |
|
455 |
static |
456 |
alterneigh(rad,i,b,g,r) /* accepts biased BGR values */ |
457 |
int rad,i; |
458 |
register int b,g,r; |
459 |
{ |
460 |
register int j,k,lo,hi,a; |
461 |
register int *p, *q; |
462 |
|
463 |
lo = i-rad; |
464 |
if (lo<-1) lo= -1; |
465 |
hi = i+rad; |
466 |
if (hi>clrtabsiz) hi=clrtabsiz; |
467 |
|
468 |
j = i+1; |
469 |
k = i-1; |
470 |
q = radpower; |
471 |
while ((j<hi) || (k>lo)) { |
472 |
a = (*(++q)); |
473 |
if (j<hi) { |
474 |
p = network[j]; |
475 |
*p -= (a*(*p - b)) / alpharadbias; |
476 |
p++; |
477 |
*p -= (a*(*p - g)) / alpharadbias; |
478 |
p++; |
479 |
*p -= (a*(*p - r)) / alpharadbias; |
480 |
j++; |
481 |
} |
482 |
if (k>lo) { |
483 |
p = network[k]; |
484 |
*p -= (a*(*p - b)) / alpharadbias; |
485 |
p++; |
486 |
*p -= (a*(*p - g)) / alpharadbias; |
487 |
p++; |
488 |
*p -= (a*(*p - r)) / alpharadbias; |
489 |
k--; |
490 |
} |
491 |
} |
492 |
} |
493 |
|
494 |
|
495 |
static |
496 |
altersingle(alpha,j,b,g,r) /* accepts biased BGR values */ |
497 |
register int alpha,j,b,g,r; |
498 |
{ |
499 |
register int *q; |
500 |
|
501 |
q = network[j]; /* alter hit neuron */ |
502 |
*q -= (alpha*(*q - b)) / initalpha; |
503 |
q++; |
504 |
*q -= (alpha*(*q - g)) / initalpha; |
505 |
q++; |
506 |
*q -= (alpha*(*q - r)) / initalpha; |
507 |
} |
508 |
|
509 |
|
510 |
static |
511 |
learn() |
512 |
{ |
513 |
register int i,j,b,g,r; |
514 |
int radius,rad,alpha,step,delta,upto; |
515 |
register unsigned char *p; |
516 |
unsigned char *lim; |
517 |
|
518 |
upto = lengthcount/(3*samplefac); |
519 |
delta = upto/ncycles; |
520 |
lim = thepicture + lengthcount; |
521 |
p = thepicture; |
522 |
alpha = initalpha; |
523 |
radius = initradius; |
524 |
rad = radius >> radiusbiasshift; |
525 |
if (rad <= 1) rad = 0; |
526 |
for (i=0; i<rad; i++) |
527 |
radpower[i] = alpha*(((rad*rad - i*i)*radbias)/(rad*rad)); |
528 |
|
529 |
if ((lengthcount%jump1) != 0) step = 3*jump1; |
530 |
else { |
531 |
if ((lengthcount%jump2) !=0) step = 3*jump2; |
532 |
else { |
533 |
if ((lengthcount%jump3) !=0) step = 3*jump3; |
534 |
else step = 3*jump4; |
535 |
} |
536 |
} |
537 |
i = 0; |
538 |
while (i < upto) { |
539 |
b = p[0] << netbiasshift; |
540 |
g = p[1] << netbiasshift; |
541 |
r = p[2] << netbiasshift; |
542 |
j = contest(b,g,r); |
543 |
|
544 |
altersingle(alpha,j,b,g,r); |
545 |
if (rad) alterneigh(rad,j,b,g,r); |
546 |
/* alter neighbours */ |
547 |
|
548 |
p += step; |
549 |
if (p >= lim) p -= lengthcount; |
550 |
|
551 |
i++; |
552 |
if (i%delta == 0) { |
553 |
alpha -= alpha / alphadec; |
554 |
radius -= radius / radiusdec; |
555 |
rad = radius >> radiusbiasshift; |
556 |
if (rad <= 1) rad = 0; |
557 |
for (j=0; j<rad; j++) |
558 |
radpower[j] = alpha*(((rad*rad - j*j)*radbias)/(rad*rad)); |
559 |
} |
560 |
} |
561 |
} |
562 |
|
563 |
static |
564 |
unbiasnet() |
565 |
{ |
566 |
int i,j; |
567 |
|
568 |
for (i=0; i<clrtabsiz; i++) { |
569 |
for (j=0; j<3; j++) |
570 |
network[i][j] >>= netbiasshift; |
571 |
network[i][3] = i; /* record colour no */ |
572 |
} |
573 |
} |
574 |
|
575 |
/* Don't do this until the network has been unbiased */ |
576 |
|
577 |
static |
578 |
cpyclrtab() |
579 |
{ |
580 |
register int i,j,k; |
581 |
|
582 |
for (j=0; j<clrtabsiz; j++) { |
583 |
k = network[j][3]; |
584 |
for (i = 0; i < 3; i++) |
585 |
clrtab[k][i] = network[j][2-i]; |
586 |
} |
587 |
} |