-
Notifications
You must be signed in to change notification settings - Fork 28
/
neuquant.py
437 lines (354 loc) · 15.1 KB
/
neuquant.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
import numpy as np
from PIL import Image
def get_cKDTree():
try:
from scipy.spatial import cKDTree
except ImportError:
cKDTree = None
return cKDTree
class NeuQuant:
""" NeuQuant(image, samplefac=10, colors=256)
samplefac should be an integer number of 1 or higher, 1
being the highest quality, but the slowest performance.
With avalue of 10, one tenth of all pixels are used during
training. This value seems a nice tradeof between speed
and quality.
colors is the amount of colors to reduce the image to. This
should best be a power of two.
See also:
http://members.ozemail.com.au/~dekker/NEUQUANT.HTML
License of the NeuQuant Neural-Net Quantization Algorithm
---------------------------------------------------------
Copyright (c) 1994 Anthony Dekker
Ported to python by Marius van Voorden in 2010
NEUQUANT Neural-Net quantization algorithm by Anthony Dekker, 1994.
See "Kohonen neural networks for optimal colour quantization"
in "network: Computation in Neural Systems" Vol. 5 (1994) pp 351-367.
for a discussion of the algorithm.
See also http://members.ozemail.com.au/~dekker/NEUQUANT.HTML
Any party obtaining a copy of these files from the author, directly or
indirectly, is granted, free of charge, a full and unrestricted irrevocable,
world-wide, paid up, royalty-free, nonexclusive right and license to deal
in this software and documentation files (the "Software"), including without
limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
and/or sell copies of the Software, and to permit persons who receive
copies from any such party to do so, with the only requirement being
that this copyright notice remain intact.
"""
NCYCLES = None # Number of learning cycles
NETSIZE = None # Number of colours used
SPECIALS = None # Number of reserved colours used
BGCOLOR = None # Reserved background colour
CUTNETSIZE = None
MAXNETPOS = None
INITRAD = None # For 256 colours, radius starts at 32
RADIUSBIASSHIFT = None
RADIUSBIAS = None
INITBIASRADIUS = None
RADIUSDEC = None # Factor of 1/30 each cycle
ALPHABIASSHIFT = None
INITALPHA = None # biased by 10 bits
GAMMA = None
BETA = None
BETAGAMMA = None
network = None # The network itself
colormap = None # The network itself
netindex = None # For network lookup - really 256
bias = None # Bias and freq arrays for learning
freq = None
pimage = None
# Four primes near 500 - assume no image has a length so large
# that it is divisible by all four primes
PRIME1 = 499
PRIME2 = 491
PRIME3 = 487
PRIME4 = 503
MAXPRIME = PRIME4
pixels = None
samplefac = None
a_s = None
def setconstants(self, samplefac, colors):
self.NCYCLES = 100 # Number of learning cycles
self.NETSIZE = colors # Number of colours used
self.SPECIALS = 3 # Number of reserved colours used
self.BGCOLOR = self.SPECIALS-1 # Reserved background colour
self.CUTNETSIZE = self.NETSIZE - self.SPECIALS
self.MAXNETPOS = self.NETSIZE - 1
self.INITRAD = self.NETSIZE//8 # For 256 colours, radius starts at 32
self.RADIUSBIASSHIFT = 6
self.RADIUSBIAS = 1 << self.RADIUSBIASSHIFT
self.INITBIASRADIUS = self.INITRAD * self.RADIUSBIAS
self.RADIUSDEC = 30 # Factor of 1/30 each cycle
self.ALPHABIASSHIFT = 10 # Alpha starts at 1
self.INITALPHA = 1 << self.ALPHABIASSHIFT # biased by 10 bits
self.GAMMA = 1024.0
self.BETA = 1.0/1024.0
self.BETAGAMMA = self.BETA * self.GAMMA
self.network = np.empty((self.NETSIZE, 3), dtype='float64') # The network itself
self.colormap = np.empty((self.NETSIZE, 4), dtype='int32') # The network itself
self.netindex = np.empty(256, dtype='int32') # For network lookup - really 256
self.bias = np.empty(self.NETSIZE, dtype='float64') # Bias and freq arrays for learning
self.freq = np.empty(self.NETSIZE, dtype='float64')
self.pixels = None
self.samplefac = samplefac
self.a_s = {}
def __init__(self, image, samplefac=10, colors=256):
# Check Numpy
if np is None:
raise RuntimeError("Need Numpy for the NeuQuant algorithm.")
# Check image
if image.size[0] * image.size[1] < NeuQuant.MAXPRIME:
raise IOError("Image is too small")
if image.mode != "RGBA":
raise IOError("Image mode should be RGBA.")
# Initialize
self.setconstants(samplefac, colors)
self.pixels = np.fromstring(image.tobytes(), np.uint32)
self.setUpArrays()
self.learn()
self.fix()
self.inxbuild()
def writeColourMap(self, rgb, outstream):
for i in range(self.NETSIZE):
bb = self.colormap[i,0];
gg = self.colormap[i,1];
rr = self.colormap[i,2];
outstream.write(rr if rgb else bb)
outstream.write(gg)
outstream.write(bb if rgb else rr)
return self.NETSIZE
def setUpArrays(self):
self.network[0,0] = 0.0 # Black
self.network[0,1] = 0.0
self.network[0,2] = 0.0
self.network[1,0] = 255.0 # White
self.network[1,1] = 255.0
self.network[1,2] = 255.0
# RESERVED self.BGCOLOR # Background
for i in range(self.SPECIALS):
self.freq[i] = 1.0 / self.NETSIZE
self.bias[i] = 0.0
for i in range(self.SPECIALS, self.NETSIZE):
p = self.network[i]
p[:] = (255.0 * (i-self.SPECIALS)) / self.CUTNETSIZE
self.freq[i] = 1.0 / self.NETSIZE
self.bias[i] = 0.0
# Omitted: setPixels
def altersingle(self, alpha, i, b, g, r):
"""Move neuron i towards biased (b,g,r) by factor alpha"""
n = self.network[i] # Alter hit neuron
n[0] -= (alpha*(n[0] - b))
n[1] -= (alpha*(n[1] - g))
n[2] -= (alpha*(n[2] - r))
def geta(self, alpha, rad):
try:
return self.a_s[(alpha, rad)]
except KeyError:
length = rad*2-1
mid = length//2
q = np.array(list(range(mid-1, -1, -1))+list(range(-1, mid)))
a = alpha*(rad*rad - q*q)/(rad*rad)
a[mid] = 0
self.a_s[(alpha, rad)] = a
return a
def alterneigh(self, alpha, rad, i, b, g, r):
if i-rad >= self.SPECIALS-1:
lo = i-rad
start = 0
else:
lo = self.SPECIALS-1
start = (self.SPECIALS-1 - (i-rad))
if i+rad <= self.NETSIZE:
hi = i+rad
end = rad*2-1
else:
hi = self.NETSIZE
end = (self.NETSIZE - (i+rad))
a = self.geta(alpha, rad)[start:end]
p = self.network[lo+1:hi]
p -= np.transpose(np.transpose(p - np.array([b, g, r])) * a)
#def contest(self, b, g, r):
# """ Search for biased BGR values
# Finds closest neuron (min dist) and updates self.freq
# finds best neuron (min dist-self.bias) and returns position
# for frequently chosen neurons, self.freq[i] is high and self.bias[i] is negative
# self.bias[i] = self.GAMMA*((1/self.NETSIZE)-self.freq[i])"""
#
# i, j = self.SPECIALS, self.NETSIZE
# dists = abs(self.network[i:j] - np.array([b,g,r])).sum(1)
# bestpos = i + np.argmin(dists)
# biasdists = dists - self.bias[i:j]
# bestbiaspos = i + np.argmin(biasdists)
# self.freq[i:j] -= self.BETA * self.freq[i:j]
# self.bias[i:j] += self.BETAGAMMA * self.freq[i:j]
# self.freq[bestpos] += self.BETA
# self.bias[bestpos] -= self.BETAGAMMA
# return bestbiaspos
def contest(self, b, g, r):
""" Search for biased BGR values
Finds closest neuron (min dist) and updates self.freq
finds best neuron (min dist-self.bias) and returns position
for frequently chosen neurons, self.freq[i] is high and self.bias[i] is negative
self.bias[i] = self.GAMMA*((1/self.NETSIZE)-self.freq[i])"""
i, j = self.SPECIALS, self.NETSIZE
dists = abs(self.network[i:j] - np.array([b,g,r])).sum(1)
bestpos = i + np.argmin(dists)
biasdists = dists - self.bias[i:j]
bestbiaspos = i + np.argmin(biasdists)
self.freq[i:j] *= (1-self.BETA)
self.bias[i:j] += self.BETAGAMMA * self.freq[i:j]
self.freq[bestpos] += self.BETA
self.bias[bestpos] -= self.BETAGAMMA
return bestbiaspos
def specialFind(self, b, g, r):
for i in range(self.SPECIALS):
n = self.network[i]
if n[0] == b and n[1] == g and n[2] == r:
return i
return -1
def learn(self):
biasRadius = self.INITBIASRADIUS
alphadec = 30 + ((self.samplefac-1)/3)
lengthcount = self.pixels.size
samplepixels = lengthcount / self.samplefac
delta = samplepixels / self.NCYCLES
alpha = self.INITALPHA
i = 0;
rad = biasRadius >> self.RADIUSBIASSHIFT
if rad <= 1:
rad = 0
print("Beginning 1D learning: samplepixels = %1.2f rad = %i" %
(samplepixels, rad) )
step = 0
pos = 0
if lengthcount%NeuQuant.PRIME1 != 0:
step = NeuQuant.PRIME1
elif lengthcount%NeuQuant.PRIME2 != 0:
step = NeuQuant.PRIME2
elif lengthcount%NeuQuant.PRIME3 != 0:
step = NeuQuant.PRIME3
else:
step = NeuQuant.PRIME4
i = 0
printed_string = ''
while i < samplepixels:
if i%100 == 99:
tmp = '\b'*len(printed_string)
printed_string = str((i+1)*100/samplepixels)+"%\n"
print(tmp + printed_string)
p = self.pixels[pos]
r = (p >> 16) & 0xff
g = (p >> 8) & 0xff
b = (p ) & 0xff
if i == 0: # Remember background colour
self.network[self.BGCOLOR] = [b, g, r]
j = self.specialFind(b, g, r)
if j < 0:
j = self.contest(b, g, r)
if j >= self.SPECIALS: # Don't learn for specials
a = (1.0 * alpha) / self.INITALPHA
self.altersingle(a, j, b, g, r)
if rad > 0:
self.alterneigh(a, rad, j, b, g, r)
pos = (pos+step)%lengthcount
i += 1
if i%delta == 0:
alpha -= alpha / alphadec
biasRadius -= biasRadius / self.RADIUSDEC
rad = biasRadius >> self.RADIUSBIASSHIFT
if rad <= 1:
rad = 0
finalAlpha = (1.0*alpha)/self.INITALPHA
print("Finished 1D learning: final alpha = %1.2f!" % finalAlpha)
def fix(self):
for i in range(self.NETSIZE):
for j in range(3):
x = int(0.5 + self.network[i,j])
x = max(0, x)
x = min(255, x)
self.colormap[i,j] = x
self.colormap[i,3] = i
def inxbuild(self):
previouscol = 0
startpos = 0
for i in range(self.NETSIZE):
p = self.colormap[i]
q = None
smallpos = i
smallval = p[1] # Index on g
# Find smallest in i..self.NETSIZE-1
for j in range(i+1, self.NETSIZE):
q = self.colormap[j]
if q[1] < smallval: # Index on g
smallpos = j
smallval = q[1] # Index on g
q = self.colormap[smallpos]
# Swap p (i) and q (smallpos) entries
if i != smallpos:
p[:],q[:] = q, p.copy()
# smallval entry is now in position i
if smallval != previouscol:
self.netindex[previouscol] = (startpos+i) >> 1
for j in range(previouscol+1, smallval):
self.netindex[j] = i
previouscol = smallval
startpos = i
self.netindex[previouscol] = (startpos+self.MAXNETPOS) >> 1
for j in range(previouscol+1, 256): # Really 256
self.netindex[j] = self.MAXNETPOS
def paletteImage(self):
""" PIL weird interface for making a paletted image: create an image which
already has the palette, and use that in Image.quantize. This function
returns this palette image. """
if self.pimage is None:
palette = []
for i in range(self.NETSIZE):
palette.extend(self.colormap[i][:3])
palette.extend([0]*(256-self.NETSIZE)*3)
# a palette image to use for quant
self.pimage = Image.new("P", (1, 1), 0)
self.pimage.putpalette(palette)
return self.pimage
def quantize(self, image):
""" Use a kdtree to quickly find the closest palette colors for the pixels """
if get_cKDTree():
return self.quantize_with_scipy(image)
else:
print('Scipy not available, falling back to slower version.')
return self.quantize_without_scipy(image)
def quantize_with_scipy(self, image):
w,h = image.size
px = np.asarray(image).copy()
px2 = px[:,:,:3].reshape((w*h,3))
cKDTree = get_cKDTree()
kdtree = cKDTree(self.colormap[:,:3],leafsize=10)
result = kdtree.query(px2)
colorindex = result[1]
print("Distance: %1.2f" % (result[0].sum()/(w*h)) )
px2[:] = self.colormap[colorindex,:3]
return Image.fromarray(px).convert("RGB").quantize(palette=self.paletteImage())
def quantize_without_scipy(self, image):
"""" This function can be used if no scipy is availabe.
It's 7 times slower though.
"""
w,h = image.size
px = np.asarray(image).copy()
memo = {}
for j in range(w):
for i in range(h):
key = (px[i,j,0],px[i,j,1],px[i,j,2])
try:
val = memo[key]
except KeyError:
val = self.convert(*key)
memo[key] = val
px[i,j,0],px[i,j,1],px[i,j,2] = val
return Image.fromarray(px).convert("RGB").quantize(palette=self.paletteImage())
def convert(self, *color):
i = self.inxsearch(*color)
return self.colormap[i,:3]
def inxsearch(self, r, g, b):
"""Search for BGR values 0..255 and return colour index"""
dists = (self.colormap[:,:3] - np.array([r,g,b]))
a= np.argmin((dists*dists).sum(1))
return a