def readtext(self, image, decoder = 'greedy', beamWidth= 5, batch_size = 1,\
workers = 0, allowlist = None, blocklist = None, detail = 1,\
paragraph = False, min_size = 20,\
contrast_ths = 0.1,adjust_contrast = 0.5, filter_ths = 0.003,\
text_threshold = 0.7, low_text = 0.4, link_threshold = 0.4,\
canvas_size = 2560, mag_ratio = 1.,\
slope_ths = 0.1, ycenter_ths = 0.5, height_ths = 0.5,\
width_ths = 0.5, add_margin = 0.1):
'''
Parameters:
image: file path or numpy-array or a byte stream object
'''
img, img_cv_grey = reformat_input(image)
horizontal_list, free_list = self.detect(img, min_size, text_threshold,\
low_text, link_threshold,\
canvas_size, mag_ratio,\
slope_ths, ycenter_ths,\
height_ths,width_ths,\
add_margin, False)
result = self.recognize(img_cv_grey, horizontal_list, free_list,\
decoder, beamWidth, batch_size,\
workers, allowlist, blocklist, detail,\
paragraph, contrast_ths, adjust_contrast,\
filter_ths, False)
return result
def detect(self, img, min_size = 20, text_threshold = 0.7, low_text = 0.4,\
link_threshold = 0.4,canvas_size = 2560, mag_ratio = 1.,\
slope_ths = 0.1, ycenter_ths = 0.5, height_ths = 0.5,\
width_ths = 0.5, add_margin = 0.1, reformat=True):
if reformat:
img, img_cv_grey = reformat_input(img)
text_box = get_textbox(self.detector, img, canvas_size, mag_ratio,\
text_threshold, link_threshold, low_text,\
False, self.device)
horizontal_list, free_list = group_text_box(text_box, slope_ths,\
ycenter_ths, height_ths,\
width_ths, add_margin)
if min_size:
horizontal_list = [
i for i in horizontal_list
if max(i[1] - i[0], i[3] - i[2]) > min_size
]
free_list = [
i for i in free_list
if max(diff([c[0] for c in i]), diff([c[1]
for c in i])) > min_size
]
return horizontal_list, free_list
def recognize(self,
img,
horizontal_list=None,
free_list=None,
reformat=True,
imgH=32):
if reformat:
img, img_cv_grey = reformat_input(img_cv_grey)
if (horizontal_list == None) and (free_list == None):
b, y_max, x_max = img.shape
ratio = x_max / y_max
max_width = int(imgH * ratio)
crop_img = cv2.resize(img, (max_width, imgH),
interpolation=Image.ANTIALIAS)
image_list = [([[0, 0], [x_max, 0], [x_max, y_max],
[0, y_max]], crop_img)]
else:
image_list, max_width = get_image_list(horizontal_list,
free_list,
img,
model_height=imgH)
result = get_text(self.recognizer, image_list)
return result
def train_model(model_builder,
images,
labels,
fold_pairs,
epochs=10,
batch_size=32,
callbacks=[],
verbose=0,
fold_verbose=0):
mean_val_score = 0
mean_test_score = 0
for i in range(len(fold_pairs)):
model = model_builder()
image_preprocessor = ImagePreprocessor()
(X_train, y_train), (X_val, y_val), (X_test, y_test) = reformat_input(
images, labels, fold_pairs[i])
X_train = image_preprocessor.fit_transform(X_train)
X_val = image_preprocessor.transform(X_val)
X_test = image_preprocessor.transform(X_test)
model.fit(X_train,
y_train,
verbose=fold_verbose,
epochs=epochs,
batch_size=batch_size,
validation_data=(X_val, y_val),
callbacks=callbacks)
y_test_predict = np.argmax(model.predict(X_test), axis=1)
y_val_predict = np.argmax(model.predict(X_val), axis=1)
val_accuracy_score = accuracy_score(y_val, y_val_predict)
test_accuracy_score = accuracy_score(y_test, y_test_predict)
mean_val_score += val_accuracy_score
mean_test_score += test_accuracy_score
if verbose >= 2:
print('===========Fold {}/{}==========='.format(
i + 1, len(fold_pairs)))
print('Validation Accuracy score: ', val_accuracy_score)
print('Test Accuracy score: ', test_accuracy_score)
if verbose == 3:
print('Test Precision score: ',
precision_score(y_test, y_test_predict, average=None))
print('Test Recall score: ',
recall_score(y_test, y_test_predict, average=None))
print('Test F1 score: ',
f1_score(y_test, y_test_predict, average=None))
if verbose >= 1:
print('\nMean validation accuracy score: ',
mean_val_score / len(fold_pairs))
print('Mean test accuracy score: ', mean_test_score / len(fold_pairs))
def recognize(self, img_cv_grey, horizontal_list=None, free_list=None,\
decoder = 'greedy', beamWidth= 5, batch_size = 1,\
workers = 0, allowlist = None, blocklist = None, detail = 1,\
paragraph = False,\
contrast_ths = 0.1,adjust_contrast = 0.5, filter_ths = 0.003,\
reformat=True):
if reformat:
img, img_cv_grey = reformat_input(img_cv_grey)
if (horizontal_list == None) and (free_list == None):
y_max, x_max = img_cv_grey.shape
ratio = x_max / y_max
max_width = int(imgH * ratio)
crop_img = cv2.resize(img_cv_grey, (max_width, imgH),
interpolation=Image.ANTIALIAS)
image_list = [([[0, 0], [x_max, 0], [x_max, y_max],
[0, y_max]], crop_img)]
else:
image_list, max_width = get_image_list(horizontal_list,
free_list,
img_cv_grey,
model_height=imgH)
if allowlist:
ignore_char = ''.join(set(self.character) - set(allowlist))
elif blocklist:
ignore_char = ''.join(set(blocklist))
else:
ignore_char = ''.join(set(self.character) - set(self.lang_char))
if self.model_lang in [
'chinese_tra', 'chinese_sim', 'japanese', 'korean'
]:
decoder = 'greedy'
result = get_text(self.character, imgH, int(max_width), self.recognizer, self.converter, image_list,\
ignore_char, decoder, beamWidth, batch_size, contrast_ths, adjust_contrast, filter_ths,\
workers, self.device)
if self.model_lang == 'arabic':
direction_mode = 'rtl'
result = [list(item) for item in result]
for item in result:
item[1] = get_display(item[1])
else:
direction_mode = 'ltr'
if paragraph:
result = get_paragraph(result, mode=direction_mode)
if detail == 0:
return [item[1] for item in result]
else:
return result
def readtext(self, image,min_size = 20,\
text_threshold = 0.7, low_text = 0.4, link_threshold = 0.4,\
canvas_size = 2560, mag_ratio = 1.,\
slope_ths = 0.1, ycenter_ths = 0.5, height_ths = 0.5,\
width_ths = 0.5, add_margin = 0.1):
'''
Parameters:
image: file path or numpy-array or a byte stream object
'''
img, img_cv_grey = reformat_input(image)
horizontal_list, free_list = self.detect(img, min_size, text_threshold,\
low_text, link_threshold,\
canvas_size, mag_ratio,\
slope_ths, ycenter_ths,\
height_ths,width_ths,\
add_margin, False)
result = self.recognize(img, horizontal_list, free_list, False)
return result, img
def recognize(self, img_cv_grey, horizontal_list=None, free_list=None,\
decoder = 'greedy', beamWidth= 5, batch_size = 1,\
workers = 0, allowlist = None, blocklist = None, detail = 1,\
rotation_info = None,\
paragraph = False,\
contrast_ths = 0.1,adjust_contrast = 0.5, filter_ths = 0.003,\
reformat=True):
if reformat:
img, img_cv_grey = reformat_input(img_cv_grey)
if allowlist:
ignore_char = ''.join(set(self.character) - set(allowlist))
elif blocklist:
ignore_char = ''.join(set(blocklist))
else:
ignore_char = ''.join(set(self.character) - set(self.lang_char))
if self.model_lang in ['chinese_tra', 'chinese_sim']:
decoder = 'greedy'
if (horizontal_list == None) and (free_list == None):
y_max, x_max = img_cv_grey.shape
horizontal_list = [[0, x_max, 0, y_max]]
free_list = []
# without gpu/parallelization, it is faster to process image one by one
if ((batch_size == 1) or (self.device == 'cpu')) and not rotation_info:
result = []
for bbox in horizontal_list:
h_list = [bbox]
f_list = []
image_list, max_width = get_image_list(h_list,
f_list,
img_cv_grey,
model_height=imgH)
result0 = get_text(self.character, imgH, int(max_width), self.recognizer, self.converter, image_list,\
ignore_char, decoder, beamWidth, batch_size, contrast_ths, adjust_contrast, filter_ths,\
workers, self.device)
result += result0
for bbox in free_list:
h_list = []
f_list = [bbox]
image_list, max_width = get_image_list(h_list,
f_list,
img_cv_grey,
model_height=imgH)
result0 = get_text(self.character, imgH, int(max_width), self.recognizer, self.converter, image_list,\
ignore_char, decoder, beamWidth, batch_size, contrast_ths, adjust_contrast, filter_ths,\
workers, self.device)
result += result0
# default mode will try to process multiple boxes at the same time
else:
image_list, max_width = get_image_list(horizontal_list,
free_list,
img_cv_grey,
model_height=imgH)
image_len = len(image_list)
if rotation_info and image_list:
image_list = make_rotated_img_list(rotation_info, image_list)
max_width = max(max_width, imgH)
result = get_text(self.character, imgH, int(max_width), self.recognizer, self.converter, image_list,\
ignore_char, decoder, beamWidth, batch_size, contrast_ths, adjust_contrast, filter_ths,\
workers, self.device)
if rotation_info and (horizontal_list + free_list):
result = set_result_with_confidence(result, image_len)
if self.model_lang == 'arabic':
direction_mode = 'rtl'
result = [list(item) for item in result]
for item in result:
item[1] = get_display(item[1])
else:
direction_mode = 'ltr'
if paragraph:
result = get_paragraph(result, mode=direction_mode)
if detail == 0:
return [item[1] for item in result]
else:
return result