DL(6)-图像分割网络Unet和Unet++详解实作

---

自从2015年，全卷积网络(FCN)诞生，图像分割在深度学习领域掀起旋风，同年稍晚Unet诞生，号称可用极少数据获取优质的结果，在数据可贵的医疗影像领域称王称霸。
2018年对Unet的改进Unet++诞生。
此外还有DeepLap、PSPNet等其他优质的图像分割网络，但它们的本质基本都是编码器-解码器(encode-decode)网络。编码器通常是一个预训练的分类网络，像 VGG、ResNet等，然后是一个解码器网络。这些架构不同的地方主要在于解码器网络。

Unet

• 网络结构
  Unet网络结构如其名呈现一个U字形，即由卷积和池化单元构成，左半边为编码器即如传统的分类网络是“下采样阶段”，右半边为解码器是“上采样阶段”，中间的灰色箭头为跳跃连接，将浅层的特征与深层的特征拼接，因为浅层通常可以抓取图像的一些简单的特征，比如边界，颜色。深层经过的卷积操作多抓取到图像的一些说不清道不明的抽象特征，将浅深同时利用起来为上上策，同时允许解码器学习在编码器池化下采样中丢失的相关特征。纵观整个网络Unet其实有点带有残差(residual)结构的思想。结构如下两幅图：
  
  
• Keras实现unet
  论文“上采样”采用的线性插值方式，keras的api即UpSampling2D，这里采用转置卷积的方式keras的api即Conv2DTranspose。论文中并没有采用padding，因此输入和输出大小并不一样，这里采用padding保持输入与输出大小一致。

def unet(input_size = (512,512,3),base_filter_num=64):
    inputs = Input(input_size)
    conv1 = Conv2D(base_filter_num, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
    conv1 = Conv2D(base_filter_num, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv1)
    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)

    conv2 = Conv2D(base_filter_num*2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
    conv2 = Conv2D(base_filter_num*2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv2)
    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)

    conv3 = Conv2D(base_filter_num*4, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
    conv3 = Conv2D(base_filter_num*4, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv3)
    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)

    conv4 = Conv2D(base_filter_num*8, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
    conv4 = Conv2D(base_filter_num*8, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv4)
    pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)

    conv5 = Conv2D(base_filter_num*16, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
    conv5 = Conv2D(base_filter_num*16, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv5)

    up6 = Conv2DTranspose(base_filter_num*8, (2, 2), strides=(2, 2), padding='same')(conv5)
    # up6 = Conv2D(base_filter_num*8, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv5))
    merge6 = concatenate([conv4,up6], axis = -1)
    conv6 = Conv2D(base_filter_num*8, 3, activation = 'relu', padding = 'same', kernel_