DL之CNN：基于CNN-RNN(GRU,2)算法(keras+tensorflow)实现不定长文本识别
目录
输出结果
实现代码
输出结果
后期更新……
实现代码
后期更新……
image_ocr代码：DL之CNN：利用CNN(keras, CTC loss, {image_ocr})算法实现OCR光学字符识别
#DL之CNN：基于CNN-RNN(GRU,2)算法(keras+tensorflow)实现不定长文本识别#Keras 的 CTC loss函数：位于 /fchollet/keras/blob/master/keras/backend/tensorflow_backend.py文件中，内容如下：import tensorflow as tffrom tensorflow.python.ops import ctc_ops as ctcdef ctc_batch_cost(y_true, y_pred, input_length, label_length):"""Runs CTC loss algorithm on each batch element.# Argumentsy_true: tensor `(samples, max_string_length)`containing the truth labels.y_pred: tensor `(samples, time_steps, num_categories)`containing the prediction, or output of the softmax.input_length: tensor `(samples, 1)` containing the sequence length foreach batch item in `y_pred`.label_length: tensor `(samples, 1)` containing the sequence length foreach batch item in `y_true`.# ReturnsTensor with shape (samples,1) containing theCTC loss of each element."""label_length = tf.to_int32(tf.squeeze(label_length))input_length = tf.to_int32(tf.squeeze(input_length))sparse_labels = tf.to_int32(ctc_label_dense_to_sparse(y_true, label_length))y_pred = tf.log(tf.transpose(y_pred, perm=[1, 0, 2]) + 1e-8)return tf.expand_dims(ctc.ctc_loss(inputs=y_pred, labels=sparse_labels, sequence_length=input_length), 1)# 不定长文本识别import osimport itertoolsimport reimport datetimeimport cairocffi as cairoimport editdistanceimport numpy as npfrom scipy import ndimageimport pylabfrom keras import backend as Kfrom keras.layers.convolutional import Conv2D, MaxPooling2Dfrom keras.layers import Input, Dense, Activation, Reshape, Lambdafrom keras.layers.merge import add, concatenatefrom keras.layers.recurrent import GRUfrom keras.models import Modelfrom keras.optimizers import SGDfrom keras.utils.data_utils import get_filefrom keras.preprocessing import imagefrom keras.callbacks import EarlyStopping,Callbackfrom keras.backend.tensorflow_backend import set_sessionimport tensorflow as tfimport matplotlib.pyplot as pltconfig = tf.ConfigProto()config.gpu_options.allow_growth=Trueset_session(tf.Session(config=config))OUTPUT_DIR = 'image_ocr'np.random.seed(55)# # 从 Keras 官方文件中 import 相关的函数# !wget /fchollet/keras/master/examples/image_ocr.pyfrom image_ocr import *#定义必要的参数：run_name = datetime.datetime.now().strftime('%Y:%m:%d:%H:%M:%S')start_epoch = 0stop_epoch = 200img_w = 128img_h = 64words_per_epoch = 16000val_split = 0.2val_words = int(words_per_epoch * (val_split))# Network parametersconv_filters = 16kernel_size = (3, 3)pool_size = 2time_dense_size = 32rnn_size = 512input_shape = (img_w, img_h, 1)# 使用这些函数以及对应参数构建生成器，生成不固定长度的验证码fdir = os.path.dirname(get_file('wordlists.tgz', origin='http://www.mythic-/datasets/wordlists.tgz', untar=True))img_gen = TextImageGenerator(monogram_file=os.path.join(fdir, 'wordlist_mono_clean.txt'),bigram_file=os.path.join(fdir, 'wordlist_bi_clean.txt'),minibatch_size=32, img_w=img_w, img_h=img_h,downsample_factor=(pool_size ** 2), val_split=words_per_epoch - val_words )#构建CNN网络act = 'relu'input_data = Input(name='the_input', shape=input_shape, dtype='float32')inner = Conv2D(conv_filters, kernel_size, padding='same', activation=act, kernel_initializer='he_normal',name='conv1')(input_data)inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max1')(inner)inner = Conv2D(conv_filters, kernel_size, padding='same', activation=act, kernel_initializer='he_normal',name='conv2')(inner)inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max2')(inner)conv_to_rnn_dims = (img_w // (pool_size ** 2), (img_h // (pool_size ** 2)) * conv_filters)inner = Reshape(target_shape=conv_to_rnn_dims, name='reshape')(inner)#减少输入尺寸到RNN：cuts down input size going into RNN: inner = Dense(time_dense_size, activation=act, name='dense1')(inner)#GRU模型：两层双向的算法# Two layers of bidirecitonal GRUs# GRU seems to work as well, if not better than LSTM:gru_1 = GRU(rnn_size, return_sequences=True, kernel_initializer='he_normal', name='gru1')(inner)gru_1b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru1_b')(inner)gru1_merged = add([gru_1, gru_1b])gru_2 = GRU(rnn_size, return_sequences=True, kernel_initializer='he_normal', name='gru2')(gru1_merged)gru_2b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru2_b')(gru1_merged)#将RNN输出转换为字符激活：transforms RNN output to character activationsinner = Dense(img_gen.get_output_size(), kernel_initializer='he_normal',name='dense2')(concatenate([gru_2, gru_2b]))y_pred = Activation('softmax', name='softmax')(inner)Model(inputs=input_data, outputs=y_pred).summary()labels = Input(name='the_labels', shape=[img_gen.absolute_max_string_len], dtype='float32')input_length = Input(name='input_length', shape=[1], dtype='int64')label_length = Input(name='label_length', shape=[1], dtype='int64')#Keras目前不支持带有额外参数的loss funcs，所以CTC loss是在lambda层中实现的# Keras doesn't currently support loss funcs with extra parameters, so CTC loss is implemented in a lambda layerloss_out = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([y_pred, labels, input_length, label_length])#clipnorm似乎加快了收敛速度：clipnorm seems to speeds up convergencesgd = SGD(lr=0.02, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)#计算损失发生在其他地方，所以使用一个哑函数来表示损失# the loss calc occurs elsewhere, so use a dummy lambda func for the pile(loss={'ctc': lambda y_true, y_pred: y_pred}, optimizer=sgd)if start_epoch > 0:weight_file = os.path.join(OUTPUT_DIR, os.path.join(run_name, 'weights%02d.h5' % (start_epoch - 1)))model.load_weights(weight_file)#捕获softmax的输出，以便在可视化过程中解码输出# captures output of softmax so we can decode the output during visualizationtest_func = K.function([input_data], [y_pred])# 反馈函数，即运行固定次数后，执行反馈函数可保存模型，并且可视化当前训练的效果viz_cb = VizCallback(run_name, test_func, img_gen.next_val())# 执行训练：model.fit_generator(generator=img_gen.next_train(), steps_per_epoch=(words_per_epoch - val_words),epochs=stop_epoch, validation_data=img_gen.next_val(), validation_steps=val_words,callbacks=[EarlyStopping(patience=10), viz_cb, img_gen], initial_epoch=start_epoch)
