forked from lllcho/CAPTCHA-breaking
-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_type1.py
180 lines (172 loc) · 6.38 KB
/
test_type1.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
# __author__ = 'lllcho'
# __date__ = '2015/8/5'
import cv2
import numpy as np
import cPickle
import codecs
import scipy.spatial.distance
from keras.regularizers import l2
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from util import *
model_path = './model/chars4800_model.d5'
dic = cPickle.load(open('model/chars4800.pkl', 'rb'))
chars = dic['c']
pys = dic['p']
py_ch = dic['pc']
ch_py = dic['cp']
nb_class = 4800
model_ch = Sequential()
model_ch.add(Convolution2D(32, 1, 4, 4, border_mode='full', activation='relu'))
model_ch.add(Convolution2D(32, 32, 4, 4, activation='relu'))
model_ch.add(MaxPooling2D(poolsize=(2, 2)))
model_ch.add(Dropout(0.25))
model_ch.add(Convolution2D(64, 32, 4, 4, border_mode='full', activation='relu'))
model_ch.add(Convolution2D(64, 64, 4, 4, activation='relu'))
model_ch.add(MaxPooling2D(poolsize=(2, 2)))
model_ch.add(Dropout(0.25))
model_ch.add(Flatten())
model_ch.add(Dense(64 * 8 * 8, 512, activation='relu'))
model_ch.add(Dropout(0.5))
model_ch.add(Dense(512, nb_class, activation='softmax'))
model_ch.load_weights(model_path)
model_ch.compile(loss='categorical_crossentropy', optimizer='adagrad')
nb_class = 408
weight_decay = 0.001
model_py = Sequential()
model_py.add(Convolution2D(32, 1, 4, 4, border_mode='full', activation='relu', W_regularizer=l2(weight_decay)))
model_py.add(Convolution2D(32, 32, 4, 4, activation='relu', W_regularizer=l2(weight_decay)))
model_py.add(MaxPooling2D(poolsize=(2, 3)))
model_py.add(Dropout(0.25))
model_py.add(Convolution2D(64, 32, 4, 4, border_mode='full', activation='relu', W_regularizer=l2(weight_decay)))
model_py.add(Convolution2D(64, 64, 4, 4, activation='relu', W_regularizer=l2(weight_decay)))
model_py.add(MaxPooling2D(poolsize=(2, 3)))
model_py.add(Dropout(0.25))
model_py.add(Flatten())
model_py.add(Dense(64 * 7 * 10, 512, activation='relu', W_regularizer=l2(weight_decay)))
model_py.add(Dropout(0.5))
model_py.add(Dense(512, nb_class, activation='softmax'))
model_py.load_weights('./model/pinyin_model.d5')
model_py.compile(loss='categorical_crossentropy', optimizer='adagrad')
comp = "type1_test1"
img_dir = './image/' + comp + '/'
fcsv = codecs.open("./result/" + comp + '.csv', 'w', 'utf-8')
fcsv.write('pic_id,content\n')
# for nb_img in range(1,20001):
# name=comp+'_' + str(nb_img) + '.jpg'
import os
names = os.listdir(img_dir)
for name in names:
imgname = img_dir + name
print imgname
img = cv2.imread(imgname, cv2.IMREAD_GRAYSCALE)
im = 255 - img
retval, im = cv2.threshold(im, 255 * 0.3, 255, cv2.THRESH_BINARY)
lines = cv2.HoughLinesP(im, 1, 0.5 * np.pi / 180, 50, maxLineGap=5, minLineLength=30)
l = im - im
for i in range(lines.shape[0]):
for x1, y1, x2, y2 in lines[i]:
cv2.line(l, (x1, y1), (x2, y2), 255, 1)
im2 = im
im = im - l
x = [11, 78, 144] * 3
y = [6, 6, 6, 46, 46, 46, 86, 86, 86]
imgs = np.zeros((9, 1, 32, 32), np.float32)
for i in range(9):
imgs[i] = im2[y[i]:y[i] + 32, x[i]:x[i] + 32]
scores1 = model_ch.predict(imgs / 255)
classes1 = np.argsort(scores1, axis=1)
result1 = []
result1_2 = []
result1_3 = []
for c in range(classes1.shape[0]):
result1.append(chars[classes1[c][-1]])
result1_2.append(chars[classes1[c][-2]])
result1_3.append(chars[classes1[c][-3]])
print(''.join(result1))
img = im2[125:157, 10:im.shape[1] - 15]
t = im[125:157, 10:im.shape[1] - 15]
t = cv2.medianBlur(t, 3)
retval, t = cv2.threshold(t, 255 * 0.5, 1, cv2.THRESH_BINARY)
s = np.sum(t[5:, :], axis=0)
s2 = s.copy()
s[s <= 4] = 0
img = img[:, 0:max(110, min(s.nonzero()[0][-1] + 10, s.shape[0]))]
idxl = list(range(0, img.shape[1] - 32, 30))
idxr = map(lambda x: img.shape[1] - x - 32, idxl)
idxl = idxl[:3]
idxr = idxr[:3]
idxs = idxl + idxr
imgs = np.zeros((len(idxs), 1, 32, 32), np.float32)
for i in range(len(idxs)):
imgs[i] = img[0:32, idxs[i]:idxs[i] + 32]
scores2 = model_ch.predict(imgs / 255)
classes2 = np.argmax(scores2, axis=1)
# ftrs2 = get_features(imgs / 255)
result2 = []
for c in classes2:
result2.append(chars[c])
print(''.join(result2))
dd = scipy.spatial.distance.cdist(scores1, scores2, get_crossentropy)
idx_up = dd.argmin(axis=0)
ddm = dd.min(axis=0)
for ii in range(len(result2)):
if result2[ii] in result1:
ddm[ii] = 0
idx_up[ii] = result1.index(result2[ii])
align_iterms = get_align_terms(idxl, idxr)
for align_iterm in align_iterms:
if ddm[align_iterm[0]] > ddm[align_iterm[1]]:
ddm[align_iterm[0]] = np.inf
else:
ddm[align_iterm[1]] = np.inf
idx_t = np.argsort(ddm)
idx_t = idx_t[:3]
idx_t = idx_t[np.asarray(idxs)[idx_t].argsort()]
res = []
for idx in idx_t:
res.append(result2[idx])
print(''.join(res))
print(idx_up[idx_t] + 1)
idx_list = list(idx_up[idx_t] + 1)
py_pos = get_py_pos(np.asarray(idxs)[idx_t], img.shape[1])
img_py = np.zeros((1, 1, 28, 90))
img_py[0, 0, :, :py_pos[1] - py_pos[0]] = img[3:-1, py_pos[0]:py_pos[1]]
classes_py = model_py.predict_classes(img_py / 255, verbose=0)
pinyin = pys[classes_py]
# print(pinyin)
words = []
words2 = []
words3 = []
words_idxs = []
for i in range(9):
if i not in idx_up[idx_t]:
words.append(result1[i])
words2.append(result1_2[i])
words3.append(result1_3[i])
words_idxs.append(i)
wordspy = map(lambda w: ch_py[w], words)
wordspy2 = map(lambda w: ch_py[w], words2)
wordspy3 = map(lambda w: ch_py[w], words3)
if pinyin in flat(wordspy):
for ii in range(len(wordspy)):
if pinyin in wordspy[ii]:
pinyin_idx = ii
break
elif pinyin in flat(wordspy2):
for ii in range(len(wordspy2)):
if pinyin in wordspy2[ii]:
pinyin_idx = ii
break
elif pinyin in flat(wordspy3):
for ii in range(len(wordspy3)):
if pinyin in wordspy3[ii]:
pinyin_idx = ii
break
else:
pinyin_idx = np.random.randint(0, len(wordspy))
idx_list.insert(py_pos[0] / 30, words_idxs[pinyin_idx] + 1)
# print(idx_list)
fcsv.write(name + ',' + ''.join(map(lambda x: str(x), idx_list)) + '\n')
fcsv.close()