使用keras框架cnn+ctc_loss識別不定長字符圖片操作

更新時間：2020年06月29日 09:46:59 作者：xinfeng2005

這篇文章主要介紹了使用keras框架cnn+ctc_loss識別不定長字符圖片操作，具有很好的參考價值，希望對大家有所幫助。一起跟隨小編過來看看吧

我就廢話不多說了，大家還是直接看代碼吧~

# -*- coding: utf-8 -*-
#keras==2.0.5
#tensorflow==1.1.0

import os,sys,string
import sys
import logging
import multiprocessing
import time
import json
import cv2
import numpy as np
from sklearn.model_selection import train_test_split

import keras
import keras.backend as K
from keras.datasets import mnist
from keras.models import *
from keras.layers import *
from keras.optimizers import *
from keras.callbacks import *
from keras import backend as K
# from keras.utils.visualize_util import plot
from visual_callbacks import AccLossPlotter
plotter = AccLossPlotter(graphs=['acc', 'loss'], save_graph=True, save_graph_path=sys.path[0])

#識別字符集
char_ocr='0123456789' #string.digits
#定義識別字符串的最大長度
seq_len=8
#識別結果集合個數(shù) 0-9
label_count=len(char_ocr)+1

def get_label(filepath):
 # print(str(os.path.split(filepath)[-1]).split('.')[0].split('_')[-1])
 lab=[]
 for num in str(os.path.split(filepath)[-1]).split('.')[0].split('_')[-1]:
 lab.append(int(char_ocr.find(num)))
 if len(lab) < seq_len:
 cur_seq_len = len(lab)
 for i in range(seq_len - cur_seq_len):
  lab.append(label_count) #
 return lab

def gen_image_data(dir=r'data\train', file_list=[]):
 dir_path = dir
 for rt, dirs, files in os.walk(dir_path): # =pathDir
 for filename in files:
  # print (filename)
  if filename.find('.') >= 0:
  (shotname, extension) = os.path.splitext(filename)
  # print shotname,extension
  if extension == '.tif': # extension == '.png' or
   file_list.append(os.path.join('%s\\%s' % (rt, filename)))
   # print (filename)

 print(len(file_list))
 index = 0
 X = []
 Y = []
 for file in file_list:

 index += 1
 # if index>1000:
 # break
 # print(file)
 img = cv2.imread(file, 0)
 # print(np.shape(img))
 # cv2.namedWindow("the window")
 # cv2.imshow("the window",img)
 img = cv2.resize(img, (150, 50), interpolation=cv2.INTER_CUBIC)
 img = cv2.transpose(img,(50,150))
 img =cv2.flip(img,1)
 # cv2.namedWindow("the window")
 # cv2.imshow("the window",img)
 # cv2.waitKey()
 img = (255 - img) / 256 # 反色處理
 X.append([img])
 Y.append(get_label(file))
 # print(get_label(file))
 # print(np.shape(X))
 # print(np.shape(X))

 # print(np.shape(X))
 X = np.transpose(X, (0, 2, 3, 1))
 X = np.array(X)
 Y = np.array(Y)
 return X,Y

# the actual loss calc occurs here despite it not being
# an internal Keras loss function

def ctc_lambda_func(args):
 y_pred, labels, input_length, label_length = args
 # the 2 is critical here since the first couple outputs of the RNN
 # tend to be garbage:
 # y_pred = y_pred[:, 2:, :] 測試感覺沒影響
 y_pred = y_pred[:, :, :]
 return K.ctc_batch_cost(labels, y_pred, input_length, label_length)

if __name__ == '__main__':
 height=150
 width=50
 input_tensor = Input((height, width, 1))
 x = input_tensor
 for i in range(3):
 x = Convolution2D(32*2**i, (3, 3), activation='relu', padding='same')(x)
 # x = Convolution2D(32*2**i, (3, 3), activation='relu')(x)
 x = MaxPooling2D(pool_size=(2, 2))(x)

 conv_shape = x.get_shape()
 # print(conv_shape)
 x = Reshape(target_shape=(int(conv_shape[1]), int(conv_shape[2] * conv_shape[3])))(x)

 x = Dense(32, activation='relu')(x)

 gru_1 = GRU(32, return_sequences=True, kernel_initializer='he_normal', name='gru1')(x)
 gru_1b = GRU(32, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru1_b')(x)
 gru1_merged = add([gru_1, gru_1b]) ###################

 gru_2 = GRU(32, return_sequences=True, kernel_initializer='he_normal', name='gru2')(gru1_merged)
 gru_2b = GRU(32, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru2_b')(
 gru1_merged)
 x = concatenate([gru_2, gru_2b]) ######################
 x = Dropout(0.25)(x)
 x = Dense(label_count, kernel_initializer='he_normal', activation='softmax')(x)
 base_model = Model(inputs=input_tensor, outputs=x)

 labels = Input(name='the_labels', shape=[seq_len], dtype='float32')
 input_length = Input(name='input_length', shape=[1], dtype='int64')
 label_length = Input(name='label_length', shape=[1], dtype='int64')
 loss_out = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([x, labels, input_length, label_length])

 model = Model(inputs=[input_tensor, labels, input_length, label_length], outputs=[loss_out])
 model.compile(loss={'ctc': lambda y_true, y_pred: y_pred}, optimizer='adadelta')
 model.summary()

 def test(base_model):
 file_list = []
 X, Y = gen_image_data(r'data\test', file_list)
 y_pred = base_model.predict(X)
 shape = y_pred[:, :, :].shape # 2:
 out = K.get_value(K.ctc_decode(y_pred[:, :, :], input_length=np.ones(shape[0]) * shape[1])[0][0])[:,
  :seq_len] # 2:
 print()
 error_count=0
 for i in range(len(X)):
  print(file_list[i])
  str_src = str(os.path.split(file_list[i])[-1]).split('.')[0].split('_')[-1]
  print(out[i])
  str_out = ''.join([str(x) for x in out[i] if x!=-1 ])
  print(str_src, str_out)
  if str_src!=str_out:
  error_count+=1
  print('################################',error_count)
  # img = cv2.imread(file_list[i])
  # cv2.imshow('image', img)
  # cv2.waitKey()

 class LossHistory(Callback):
 def on_train_begin(self, logs={}):
  self.losses = []

 def on_epoch_end(self, epoch, logs=None):
  model.save_weights('model_1018.w')
  base_model.save_weights('base_model_1018.w')
  test(base_model)

 def on_batch_end(self, batch, logs={}):
  self.losses.append(logs.get('loss'))


 # checkpointer = ModelCheckpoint(filepath="keras_seq2seq_1018.hdf5", verbose=1, save_best_only=True, )
 history = LossHistory()

 # base_model.load_weights('base_model_1018.w')
 # model.load_weights('model_1018.w')

 X,Y=gen_image_data()
 maxin=4900
 subseq_size = 100
 batch_size=10
 result=model.fit([X[:maxin], Y[:maxin], np.array(np.ones(len(X))*int(conv_shape[1]))[:maxin], np.array(np.ones(len(X))*seq_len)[:maxin]], Y[:maxin],
   batch_size=20,
   epochs=1000,
   callbacks=[history, plotter, EarlyStopping(patience=10)], #checkpointer, history,
   validation_data=([X[maxin:], Y[maxin:], np.array(np.ones(len(X))*int(conv_shape[1]))[maxin:], np.array(np.ones(len(X))*seq_len)[maxin:]], Y[maxin:]),
   )

 test(base_model)

 K.clear_session()

補充知識：日常填坑之keras.backend.ctc_batch_cost參數(shù)問題

InvalidArgumentError sequence_length(0) <=30錯誤

下面的代碼是在網(wǎng)上絕大多數(shù)文章給出的關于k.ctc_batch_cost()函數(shù)的使用代碼

def ctc_lambda_func(args):
 y_pred, labels, input_length, label_length = args
 # the 2 is critical here since the first couple outputs of the RNN
 # tend to be garbage: 
 y_pred = y_pred[:, 2:, :]
 return K.ctc_batch_cost(labels, y_pred, input_length, label_length)

可以注意到有一句：y_pred = y_pred[:, 2:, :]，這里把y_pred 的第二維數(shù)據(jù)去掉了兩列，說人話：把送進lstm序列的step減了2步。后來偶然在一篇文章中有提到說這里之所以減2是因為在將feature送入keras的lstm時自動少了2維，所以這里就寫成這樣了。估計是之前老版本的bug，現(xiàn)在的新版本已經(jīng)修復了。如果依然按照上面的寫法，會得到如下錯誤：

InvalidArgumentError sequence_length(0) <=30

'<='后面的數(shù)值 = 你cnn最后的輸出維度 - 2。這個錯誤我找了很久，一直不明白30哪里來的，后來一行行的檢查代碼是發(fā)現(xiàn)了這里很可疑，于是改成如下形式錯誤解決。

def ctc_lambda_func(args):
 y_pred, labels, input_length, label_length = args 
 return K.ctc_batch_cost(labels, y_pred, input_length, label_length)

訓練時出現(xiàn)ctc_loss_calculator.cc:144] No valid path found或loss: inf錯誤

熟悉CTC算法的話，這個提示應該是ctc沒找到有效路徑。既然是沒找到有效路徑，那肯定是label和input之間哪個地方又出問題了！和input相關的錯誤已經(jīng)解決了，那么肯定就是label的問題了。再看ctc_batch_cost的四個參數(shù)，labels和label_length這兩個地方有可疑。對于ctc_batch_cost()的參數(shù)，labels需要one-hot編碼，形狀：[batch, max_labelLength]，其中max_labelLength指預測的最大字符長度；label_length就是每個label中的字符長度了，受之前tf.ctc_loss的影響把這里都設置成了最大長度，所以報錯。

對于參數(shù)labels而言，max_labelLength是能預測的最大字符長度。這個值與送lstm的featue的第二維，即特征序列的max_step有關，表面上看只要max_labelLength<max_step即可，但是如果小的不多依然會出現(xiàn)上述錯誤。至于到底要小多少，還得從ctc算法里找，由于ctc算法在標簽中的每個字符后都加了一個空格，所以應該把這個長度考慮進去，所以有 max_labelLength < max_step//2。沒仔細研究keras里ctc_batch_cost()函數(shù)的實現(xiàn)細節(jié)，上面是我的猜測。如果有很明確的答案，還請麻煩告訴我一聲，謝了先！

錯誤代碼：

batch_label_length = np.ones(batch_size) * max_labelLength

正確打開方式：

batch_x, batch_y = [], []
batch_input_length = np.ones(batch_size) * (max_img_weigth//8)
batch_label_length = []
for j in range(i, i + batch_size):
 x, y = self.get_img_data(index_all[j])
 batch_x.append(x)
 batch_y.append(y)
 batch_label_length.append(self.label_length[j])

最后附一張我的crnn的模型圖：