ResNet

深度学习神经网络特征提取（一）

网络结构

要构建ResNet特征提取网络，我们只需要了解两个方面：

• ResNet网络层数：Resnet50,ResNet101,ResNet152
• ResNet网络基础组成：Conv_Block,Identity_Block

在本文中，根据实际的项目要求，构建了RseNet101网络。

构建Conv_Block模块

针对Conv_Block模块，我们首先要了解这个模块的具体结构，Conv_Block从输入开始分两支分别进行特征提取，以一次卷积、一次归一化、一次ReLu激活函数的形式连接三次，并在第一次的卷积层步长为2进行降维，另一分支只进行一次步长为2的卷积、一次归一化，然后将两个分支进行连接，再一次激活函数层完成一次Conv_Block，结构如下图：

    def Conv_block(input_feature,kernel_size,filters,strides = (2,2)):
        filter1,filter2,filter3 = filters
        #first line conv
        x = layers.Conv2D(filter1,(1,1),strides = strides, use_bias=True)(input_feature)
        x = layers.BatchNormalization()(x)
        x = layers.Activation('relu')(x)

        x = layers.Conv2D(filter2,(kernel_size,kernel_size),padding = 'same', use_bias=True)(x)
        x = layers.BatchNormalization()(x)
        x = layers.Activation('relu')(x)

        x = layers.Conv2D(filter3,(1,1), use_bias=True)(x)
        x = layers.BatchNormalization()(x)

        #parallel conv
        shortcut = layers.Conv2D(filter3,(1,1), strides = strides, use_bias=True)(input_feature)
        shortcut = layers.BatchNormalization()(shortcut)

        # add the parallel conv

        x = layers.Add()([x,shortcut])
        x = layers.Activation('relu')(x)
        return x

构建Identity_Block模块

Identity_Block模块不同于Conv_Block模块，Identity_Block模块只对特征进行提取，即只进行深度的堆叠不行进降维，所以在结构上与Conv_Block相似——同样是双分支结构，在另一分支上不进行操作，只是将输入与另一分支的结果进行叠加，具体结构如下：

代码如下：

    def identity_block(input_feature,kernel_size,filters):
        filter1,filter2,filter3 = filters

        x = layers.Conv2D(filter1,(1,1), use_bias=True)(input_feature)
        x = layers.BatchNormalization()(x)
        x = layers.Activation('relu')(x)

        x = layers.Conv2D(filter2,(kernel_size,kernel_size), padding = 'same', use_bias=True)(x)
        x = layers.BatchNormalization()(x)
        x = layers.Activation('relu')(x)

        x = layers.Conv2D(filter3,(1,1), use_bias=True)(x)
        x = layers.BatchNormalization()(x)

        x = layers.Add()([x,input_feature])
        x = layers.Activation('relu')(x)
        return x

构建ResNet101网络

构建ResNet101网络的具体形式如上图所示。

• 由于第一层7x7卷积，所以在卷积前加入ZeroPadding2D，然后如同图中所示，归一化接一层激活和池化
• 第二层一次Conv_block、两次identity_block
• 第三层一次Conv_block、三次identity_block
• 第四层一次Conv_block、22次identity_block
• 第五层一次Conv_block、两次identity_block

在第五层之后本文接上一层全局池化以及两层全连接进行分类，但是ResNset网络在第五层基本就结束了，后续根据个人的需求进行修改即可，本文只是做了一个图片的分类的实例，接上后续的全连接分类。
代码如下：

    def get_resnet_model(input_shape,classes):
        input_image = Input(shape = input_shape)
        x = layers.ZeroPadding2D((3,3))(input_image)
        x = layers.Conv2D(64,(7,7),strides=(2,2),  use_bias=True)(x)
        x = layers.BatchNormalization()(x)
        x = layers.Activation('relu')(x)
        c1 = x = layers.MaxPooling2D((3,3),strides=(2, 2), padding='same')(x)

        x = Conv_block(x, 3, [64,64,256], strides=(1,1))
        x = identity_block(x,3,[64,64,256])
        c2 = x = identity_block(x,3,[64,64,256])
        x = Conv_block(x,3,[128,128,512])
        x = identity_block(x,3,[128,128,512])
        x = identity_block(x,3,[128,128,512])
        c3 = x =identity_block(x,3,[128,128,512])
        x = Conv_block(x,3,[256,256,1024])
        for i in range(22):
            x = identity_block(x,3,[256,256,1024])
        c4 = x
        x = Conv_block(x,3,[512,512,2048])
        x = identity_block(x,3,[512,512,2048])
        c5 = x = identity_block(x,3,[512,512,2048])
        x = layers.GlobalAveragePooling2D()(x)
        x = layers.Dropout(0.5)(x)
        x = layers.Dense(1024,activation = 'relu')(x)
        x = layers.Dropout(0.5)(x)
        output = layers.Dense(classes,activation = 'softmax')(x)
        model = models.Model(input_image,output)
        return model