def main():
c.clear_temp()
img = dataset_manager.get_training_image()
recognizer = captcha_recognizer.CaptchaRecognizer()
mpimg.imsave(c.temp_path('00.origin.png'), img)
# 1
img_01 = time_func(
'remove_noise_with_hsv',
lambda: recognizer.remove_noise_with_hsv(img)
)
mpimg.imsave(c.temp_path('01.hsv.png'), img_01, cmap=cm_greys)
# 2
img_02 = time_func(
'remove_noise_with_neighbors',
lambda: repeat(recognizer.remove_noise_with_neighbors, 2)(img_01)
)
mpimg.imsave(c.temp_path('02.neighbor.png'), img_02, cmap=cm_greys)
img_03a = time_func(
'skeletonize',
lambda: morph.skeletonize(img_02)
)
mpimg.imsave(c.temp_path('03a.skeleton.png'), img_03a, cmap=cm_greys)
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
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
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
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
def test_recognize_http(show_img=False, num=1, reconstruct=False, force_partition=True):
time_start = time.time()
provider = BilibiliCaptchaProvider()
recognizer = CaptchaRecognizer()
fail = 0
right_strong = 0
right_weak = 0
wrong_strong = 0
wrong_weak = 0
for i in range(num):
image = time_func("fetch" if num == 1 else None, lambda: provider.fetch())
if show_img and num == 1:
show_image(image)
if num == 1:
success, seq, weak_confidence = recognizer.recognize(
image, save_intermediate=True, verbose=True, reconstruct=reconstruct, force_partition=force_partition
)
else:
if i == 0:
success, seq, weak_confidence = recognizer.recognize(
image,
save_intermediate=False,
verbose=False,
reconstruct=reconstruct,
force_partition=force_partition,
)
else:
success, seq, weak_confidence = recognizer.recognize(
image, save_intermediate=False, verbose=False, reconstruct=False, force_partition=force_partition
)
if success:
print(seq)
result = time_func("verify" if num == 1 else None, lambda: provider.verify(seq))
if num == 1:
print("Recognized seq is {}".format(result))
if result:
if weak_confidence:
right_weak += 1
else:
right_strong += 1
else:
if weak_confidence:
wrong_weak += 1
else:
wrong_strong += 1
else:
fail += 1
right_total = right_strong + right_weak
wrong_total = wrong_strong + wrong_weak
print("Fail: ", fail)
print("Right weak: ", right_weak)
print("Right strong: ", right_strong)
print("Right total: ", right_total)
print("Wrong weak: ", wrong_weak)
print("Wrong strong: ", wrong_strong)
print("Wrong total: ", wrong_total)
print("Total success rate: ", (right_weak + right_strong) / num)
print("Success rate when confident: ", (right_strong + right_weak) / (num - fail) if num - fail > 0 else 0)
print(
"Success rate when strongly confident: ",
right_strong / (right_strong + wrong_strong) if right_strong + wrong_strong > 0 else 0,
)
print(
"Success rate when weakly confident: ",
right_weak / (right_weak + wrong_weak) if right_weak + wrong_weak > 0 else 0,
)
time_end = time.time()
print("Time used to test recognize http is: ", time_end - time_start)
def test_recognize_http(show_img=False,
num=1,
reconstruct=False,
force_partition=True):
time_start = time.time()
provider = BilibiliCaptchaProvider()
recognizer = CaptchaRecognizer()
fail = 0
right_strong = 0
right_weak = 0
wrong_strong = 0
wrong_weak = 0
for i in range(num):
image = time_func('fetch' if num == 1 else None,
lambda: provider.fetch())
if show_img and num == 1:
show_image(image)
if num == 1:
success, seq, weak_confidence = recognizer.recognize(
image,
save_intermediate=True,
verbose=True,
reconstruct=reconstruct,
force_partition=force_partition)
else:
if i == 0:
success, seq, weak_confidence = recognizer.recognize(
image,
save_intermediate=False,
verbose=False,
reconstruct=reconstruct,
force_partition=force_partition)
else:
success, seq, weak_confidence = recognizer.recognize(
image,
save_intermediate=False,
verbose=False,
reconstruct=False,
force_partition=force_partition)
if success:
print(seq)
result = time_func('verify' if num == 1 else None,
lambda: provider.verify(seq))
if num == 1:
print('Recognized seq is {}'.format(result))
if result:
if weak_confidence:
right_weak += 1
else:
right_strong += 1
else:
if weak_confidence:
wrong_weak += 1
else:
wrong_strong += 1
else:
fail += 1
right_total = right_strong + right_weak
wrong_total = wrong_strong + wrong_weak
print('Fail: ', fail)
print('Right weak: ', right_weak)
print('Right strong: ', right_strong)
print('Right total: ', right_total)
print('Wrong weak: ', wrong_weak)
print('Wrong strong: ', wrong_strong)
print('Wrong total: ', wrong_total)
print('Total success rate: ', (right_weak + right_strong) / num)
print('Success rate when confident: ',
(right_strong + right_weak) / (num - fail) if num - fail > 0 else 0)
print(
'Success rate when strongly confident: ',
right_strong / (right_strong + wrong_strong)
if right_strong + wrong_strong > 0 else 0)
print(
'Success rate when weakly confident: ', right_weak /
(right_weak + wrong_weak) if right_weak + wrong_weak > 0 else 0)
time_end = time.time()
print('Time used to test recognize http is: ', time_end - time_start)
def partition(self, img, save_intermediate=False, verbose=False):
if save_intermediate:
mpimg.imsave(c.temp_path('00.origin.png'), img)
# 1
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)
# 2
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)
# 3
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)
# Arrange the segments from left to right
xmin_arr = np.array([s[1].start for s in object_slices])
sort_index = xmin_arr.argsort()
char_images = []
# noinspection PyTypeChecker
for i in 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)
# Check if segmentation was successful
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
if verbose:
print('Warning: partition failed!')
return None
