forked from GCC15/bilibili-captcha
/
captcha_recognizer.py
229 lines (207 loc) · 8.83 KB
/
captcha_recognizer.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
# Handle image processing before handing over to captcha learner
import matplotlib.colors as colors
import matplotlib.image as mpimg
import numpy as np
from scipy import ndimage
import config as c
from helper import time_func, cm_greys, repeat, sort_by_occurrence
from captcha_provider import BilibiliCaptchaProvider
import captcha_learn
class CaptchaRecognizer:
def __init__(self, captcha_provider=BilibiliCaptchaProvider(),
h_tol=6 / 360,
s_tol=36 / 100,
v_tol=40 / 100):
self.character_num = captcha_provider.seq_length
# Three parameters to be used in remove_noise_with_hsv
self.h_tolerance = h_tol
self.s_tolerance = s_tol
self.v_tolerance = v_tol
# parameters to be used in remove_noise_with_neighbors
self.neighbor_low = 0
self.neighbor_high = 7
self.neighbor_ratio = 1.3
# Four parameters to be used in partition
self.char_width_min = 5
self.char_width_max = 30
self.char_height_min = 10
self.char_height_max = 30
# Try to partition a CAPTCHA into each char image
def partition(self, img, save_intermediate=False, verbose=False,
force_partition=True):
weak_confidence = 0
if save_intermediate:
mpimg.imsave(c.temp_path('00.origin.png'), img)
# step 1
# remove noise with hsv
img_01 = time_func(
'remove_noise_with_hsv' if verbose else None,
lambda: self.remove_noise_with_hsv(img)
)
if save_intermediate:
mpimg.imsave(c.temp_path('01.hsv.png'), img_01, cmap=cm_greys)
# step 2
# remove noise with neighbors
img_02 = time_func(
'remove_noise_with_neighbors' if verbose else None,
lambda: repeat(self.remove_noise_with_neighbors, 2)(img_01)
)
if save_intermediate:
mpimg.imsave(c.temp_path('02.neighbor.png'), img_02, cmap=cm_greys)
# step 3
# partition stage 1
labels, object_slices = time_func(
'segment_with_label' if verbose else None,
lambda: self.segment_with_label(img_02)
)
if verbose:
print('{} connected components found'.format(len(object_slices)))
if save_intermediate:
mpimg.imsave(c.temp_path('03.00000.png'), labels)
# step 4
# Arrange the segments from left to right and probably partition stage 2
xmin_arr = np.array([s[1].start for s in object_slices])
sort_index = xmin_arr.argsort()
char_images = []
for i in list(sort_index):
char_image = img_02.copy()
char_image[labels != i + 1] = 0
char_image = char_image[object_slices[i]]
char_images.append(char_image)
if force_partition and len(char_images) == self.character_num - 1:
weak_confidence = 1
char_images = self.force_partition(char_images)
# step 5
# Check if segmentation was successful and get characters
if len(char_images) == self.character_num:
shapes = np.array(list(map(np.shape, char_images)))
heights, widths = shapes[:, 0], shapes[:, 1]
if verbose:
print('Heights {}'.format(heights))
print('Widths {}'.format(widths))
# noinspection PyTypeChecker
if (np.all(heights >= self.char_height_min) and
np.all(heights <= self.char_height_max) and
np.all(widths >= self.char_width_min) and
np.all(widths <= self.char_width_max)):
if save_intermediate:
for i in range(len(char_images)):
mpimg.imsave(
c.temp_path('03.char.{}.png'.format(i + 1)),
char_images[i], cmap=cm_greys)
return char_images, weak_confidence
if verbose:
print('Warning: partition failed!')
return None, weak_confidence
# Recognize the captcha
def recognize(self, img, save_intermediate=False, verbose=False,
reconstruct=False, force_partition=True):
seq = []
char_images, weak_confidence = self.partition(img, save_intermediate,
verbose,force_partition)
if reconstruct:
captcha_learn.reconstruct_model()
if char_images is not None and len(char_images) == self.character_num:
success = True
def predict():
nonlocal seq
for i in range(len(char_images)):
seq.append(captcha_learn.predict(char_images[i]))
time_func('predict' if verbose else None, predict)
seq = ''.join(seq)
else:
success = False
return success, seq, weak_confidence
# Convert to a grayscale image using HSV
def remove_noise_with_hsv(self, img):
# Use number of occurrences to find the standard h, s, v
# Convert to int so we can sort the colors
# noinspection PyTypeChecker
img_int = np.dot(np.rint(img * 255), np.power(256, np.arange(3)))
color_array = sort_by_occurrence(img_int.flatten())
# standard color is the 2nd most frequent color
std_color = color_array[1]
std_b, mod = divmod(std_color, 256 ** 2)
std_g, std_r = divmod(mod, 256)
# noinspection PyTypeChecker
std_h, std_s, std_v = colors.rgb_to_hsv(
np.array([std_r, std_g, std_b]) / 255
)
# print(std_h * 360, std_s * 100, std_v * 100)
height, width, _ = img.shape
img_hsv = colors.rgb_to_hsv(img)
h, s, v = img_hsv[:, :, 0], img_hsv[:, :, 1], img_hsv[:, :, 2]
h_mask = np.abs(h - std_h) > self.h_tolerance
s_mask = np.abs(s - std_s) > self.s_tolerance
delta_v = np.abs(v - std_v)
v_mask = delta_v > self.v_tolerance
hsv_mask = np.logical_or(
np.logical_or(
h_mask, s_mask
), v_mask
)
new_img = 1 - delta_v
new_img[hsv_mask] = 0
# Three types of grayscale colors in new_img:
# Type A: 1. Outside noise, or inside point.
# Type B: between 0 and 1. Outside noise, or contour point.
# Type C: 0. Inside noise, or background.
return new_img
# Adding and removing pixels on a grayscale image
def remove_noise_with_neighbors(self, img, ):
height, width = img.shape
pad_shape = height + 2, width + 2
img_pad_sum = np.zeros(pad_shape)
img_pad_a = np.zeros(pad_shape)
img_pad_b = np.zeros(pad_shape)
neighbors = [-1, 0, 1]
# Add padding in a vectorized manner
for dy in neighbors:
for dx in neighbors:
if dy == 0 and dx == 0:
continue
s = (slice(dy + 1, dy - 1 if dy - 1 else None),
slice(dx + 1, dx - 1 if dx - 1 else None))
img_pad_sum[s] += img
img_pad_a[s] += img == 1
img_pad_b[s] += np.logical_and(img > 0, img < 1)
# Remove padding
s = [slice(1, -1)] * 2
img_pad_sum = img_pad_sum[s]
img_pad_a = img_pad_a[s]
img_pad_b = img_pad_b[s]
new_img = img.copy()
mask = np.logical_and(
img == 0,
img_pad_a + img_pad_b >= self.neighbor_high
)
new_img[mask] = img_pad_sum[mask] / 8
new_img[img * self.neighbor_ratio > img_pad_sum] = 0
new_img[img_pad_a <= self.neighbor_low] = 0
return new_img
# segment a grayscale image with labels
def segment_with_label(self, img):
# Next-nearest neighbors
struct_nnn = np.ones((3, 3), dtype=int)
labels, _ = ndimage.label(img, structure=struct_nnn)
# np.savetxt(c.temp_path('labels.txt'), labels, fmt='%d')
object_slices = ndimage.find_objects(labels)
return labels, object_slices
# force a image that is separated into four parts to be further separated
def force_partition(self, char_images):
widths = []
for image in char_images:
widths.append(image.shape[1])
# The part with the largest width needs to be separate further
target_index = np.argsort(widths)[-1]
target_img = char_images[target_index]
del char_images[target_index]
width = target_img.shape[1]
if width % 2 == 1:
char_images.insert(target_index, target_img[:, 0:(width + 1) / 2])
char_images.insert(target_index + 1,
target_img[:, (width - 1) / 2:])
else:
char_images.insert(target_index, target_img[:, 0:width / 2 + 1])
char_images.insert(target_index + 1, target_img[:, width / 2 - 1:])
return char_images