def detect_functions(symbols, original_img):
sub_symbol_cnt = 3
for j in range(2):
stride = sub_symbol_cnt - j
length = len(symbols)
i = 0
t_symbols = []
locations = [x['location'] for x in symbols]
while i < length - stride + 1:
location = tools.join_locations(locations[i:i + stride])
segment_img = tools.extract_img(location, original_img)
segment = {
'src_img': segment_img,
'start_index': i,
'end_index': i + 3,
'location': location
}
t_symbols.append(segment)
i = i + 1
# print([x['start_index'] for x in t_symbols])
img_to_predict = [x['src_img'] for x in t_symbols]
# 识别字符 这里每次都需要calling model，可以进一步优化
predict_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": tools.normalize_matrix_value(img_to_predict)},
shuffle=False)
predictions = tools.cnn_symbol_classifier.predict(
input_fn=predict_input_fn)
characters = []
for i, p in enumerate(predictions):
candidates = tools.get_candidates(p['probabilities'])
characters.append(candidates[0])
# print(characters)
length = len(t_symbols)
i = 0
shift = 0
while i < length:
# 如果识别结果是函数，则替换为一个整体
if characters[i]['symbol'] in config.FUNCTION and characters[i][
'probability'] > 0.9:
start_index = t_symbols[i]['start_index']
end_index = t_symbols[i]['end_index']
# print('detect function',i,start_index,end_index,shift)
start_index = start_index - shift
end_index = end_index - shift
for j in range(end_index - start_index):
symbols.remove(symbols[start_index])
symbols.insert(start_index, t_symbols[i])
shift = shift + end_index - start_index - 1
if i + 2 < length:
i = i + 2
elif i + 1 < length:
i = i + 1
i = i + 1
return symbols
def detect_uncontinous_symbols(symbols, original_img):
# 先对symbols垂直投影
# 然后垂直投影
for projection_type in range(1):
projection = tools.get_projection(symbols, projection_type)
# print('projection',projection)
# 根据投影对symbols进行分割
# 对于每一个分割，获得在这个分割里面的symbols
locations = [x['location'] for x in symbols]
end_index = 0
for line_segment in projection:
# 确定属于line_segment的symbols
start_index = end_index
for end_index in range(start_index, len(locations)):
x11, x12 = locations[end_index][projection_type], locations[
end_index][projection_type] + locations[end_index][
projection_type + 2]
x21, x22 = line_segment[0], line_segment[1]
if (x11 >= x21 and x12 <= x22):
end_index += 1
else:
break
# 截取characters[start_index:end_index],如果长度大于1，继续递归调用sort_characters排序
location_segment = locations[start_index:end_index]
symbol_segment = symbols[start_index:end_index]
sub_symbol = [x['src_img'] for x in symbol_segment]
# print('切割后的location', location_segment)
# print(symbol_segment)
# 如果一个分割的长度大于1，小于3，就送到分类器进行识别
if (len(location_segment) > 1 and len(location_segment) < 4):
location = tools.join_locations(location_segment)
# 从原图提取待识别的图片
extracted_img = tools.extract_img(location, original_img)
# 识别字符 这里每次都需要calling model，可以进一步优化
predict_input_fn = tf.estimator.inputs.numpy_input_fn(
x={
"x":
tools.normalize_matrix_value([extracted_img] +
sub_symbol)
},
shuffle=False)
predictions = tools.cnn_symbol_classifier.predict(
input_fn=predict_input_fn)
characters = []
for i, p in enumerate(predictions):
# print(p['classes'],FILELIST[p['classes']])
candidates = tools.get_candidates(p['probabilities'])
characters.append({'candidates': candidates})
# print('detect uncontinuous symbols',characters)
# 如果识别出候选的匹配字符中存在非连续字符，即作为一个整体
for candidate in characters[0]['candidates']:
recognized_symbol = candidate['symbol']
probability = candidate['probability']
# 如果是非黏连字符,且有半成把握认为是正确的,则合并它们作为一个整体，返回合并后的symbols
if recognized_symbol in config.UNCONTINOUS_SYMBOLS and probability > 0.5:
# print('yesssssss',characters[2]['candidates'][0]['symbol'],characters[1]['candidates'][0]['symbol'].isdigit(),characters[3]['candidates'][0]['symbol'].isdigit())
# 除号必须有三个字符构成，还要判断子字符包不包含数字，如果包含数字，则不能合并为整体
if recognized_symbol == 'div' and len(characters) == 4 and \
characters[2]['candidates'][0]['symbol'] in ['-',',','point'] and \
characters[1]['candidates'][0]['symbol'].isdigit() == False and \
characters[3]['candidates'][0]['symbol'].isdigit() == False:
joined_symbol = {
'location': location,
'src_img': extracted_img
}
for i in range(end_index - start_index):
symbols.remove(symbols[start_index])
locations.remove(locations[start_index])
symbols.insert(start_index, joined_symbol)
locations.insert(start_index, location)
end_index = start_index + 1
break
# 等于号的两个子字符必须是-
elif recognized_symbol == '=' and len(characters) == 3 and \
characters[2]['candidates'][0]['symbol'] in ['-',',','point'] and \
characters[1]['candidates'][0]['symbol'] in ['-',',','point']:
joined_symbol = {
'location': location,
'src_img': extracted_img
}
for i in range(end_index - start_index):
symbols.remove(symbols[start_index])
locations.remove(locations[start_index])
symbols.insert(start_index, joined_symbol)
locations.insert(start_index, location)
end_index = start_index + 1
break
# 等于号的两个子字符必须是-
elif recognized_symbol == 'rightarrow' and len(characters) == 3 and \
characters[2]['candidates'][0]['symbol'] in [')', '>'] and \
characters[1]['candidates'][0]['symbol'] in ['-', ',', 'point']:
joined_symbol = {
'location': location,
'src_img': extracted_img
}
for i in range(end_index - start_index):
symbols.remove(symbols[start_index])
locations.remove(locations[start_index])
symbols.insert(start_index, joined_symbol)
locations.insert(start_index, location)
end_index = start_index + 1
break
return symbols