CRNN+EAST实现银行卡号定位与识别

源码：https://github.com/ShawnHXH/BankCard-Recognizer

实现工具：Python 3.6， Win10， Keras（backend is TensorFlow）

 

CRNN：

需求分析：

　　1. 银行卡号的长度大小并不是固定不变的，有的有20个字符，有的只有19个。所以模型要能够识别不定长度的卡号；

      2. 模型的输入是图像，输出是文本，故模型既需要涉及CNN也需要涉及到RNN，故称为CRNN。

模型选取：

　　1. 不定长度的识别，目前多流行采用CTC作为损失函数；

　　2. CNN则选择采用了VGG， RNN可以使用双向LSTM（BLSTM）或GRU；

模型预览：

　　1. CNN部分：

　　　　 def PatternUnits(inputs, index, activation="relu"):
            inputs = BatchNormalization(name="BN_%d" % index)(inputs)
            inputs = Activation(activation, name="Relu_%d" % index)(inputs)

            return inputs

　　　　 initializer = initializers.he_normal()
　　　　 inputs = Input(shape=(img_height, img_width, 1), name='img_inputs')
        x = Conv2D(64, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_1')(inputs)
        x = PatternUnits(x, 1)
        x = MaxPooling2D(strides=2, name='Maxpool_1')(x)
        x = Conv2D(128, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_2')(x)
        x = PatternUnits(x, 2)
        x = MaxPooling2D(strides=2, name='Maxpool_2')(x)

        x = Conv2D(256, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_3')(x)
        x = PatternUnits(x, 3)
        x = Conv2D(256, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_4')(x)
        x = PatternUnits(x, 4)
        x = MaxPooling2D(pool_size=(2, 1), strides=(2, 1), name='Maxpool_3')(x)

        x = Conv2D(512, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_5')(x)
        x = PatternUnits(x, 5)
        x = Conv2D(512, (3, 3), padding="same", kernel_initializer=initializer, name='Conv2d_6')(x)
        x = PatternUnits(x, 6)
        x = MaxPooling2D(pool_size=(2, 1), strides=(2, 1), name='Maxpool_4')(x)

        x = Conv2D(512, (2, 2), padding='same', activation='relu', kernel_initializer=initializer, name='Conv2d_7')(x)
        x = PatternUnits(x, 7)
        conv_output = MaxPooling2D(pool_size=(2, 1), name="Conv_output")(x)
        x = Permute((2, 3, 1), name='Permute')(conv_output)

　　2. RNN部分（使用BLSTM）：

rnn_input = TimeDistributed(Flatten(), name='Flatten_by_time')(x)
y = Bidirectional(LSTM(256, kernel_initializer=initializer, return_sequences=True), merge_mode='sum', name='LSTM_1')(rnn_input)
y = BatchNormalization(name='BN_8')(y)
y = Bidirectional(LSTM(256, kernel_initializer=initializer, return_sequences=True), name='LSTM_2')(y)
y_pred = Dense(num_classes, activation='softmax', name='y_pred')(y)

　　3. CTC损失函数：

def ctc_loss_layer(args):
    """
    y_true: True label.
    y_pred: Predict label.
    pred_length: Predict label length.
    label_length: True label length.
    :param args: (y_true, y_pred, pred_length, label_length).
    :return: batch_cost with shape (batch_size, 1).
    """

    y_true, y_pred, pred_length, label_length = args
    batch_cost = K.ctc_batch_cost(y_true, y_pred, pred_length, label_length)
    return batch_cost        

y_true = Input(shape=[max_label_length], name='y_true')
y_pred_length = Input(shape=[1], name='y_pred_length')
y_true_length = Input(shape=[1], name='y_true_length')
ctc_loss_output = Lambda(ctc_loss_layer, output_shape=(1,), name='ctc_loss_output')([y_true, y_pred, y_pred_length, y_true_length])

 

EAST：

当下最热的图像文本定位算法莫属CTPN，其次还有Faster RCNN， Seg-Link，Mask RCNN，EAST等等。

详情见：https://github.com/huoyijie/AdvancedEAST