【656】SegNet 相关说明

 

代码：

from keras import Model, layers
from keras.layers import Input, Conv2D, BatchNormalization, Activation, Reshape, MaxPooling2D, UpSampling2D

def segnet(pretrained_weights=None, input_size=(512, 512, 3), classNum=2, learning_rate=1e-5):
    inputs = Input(input_size)
    #encode
    #第一层 64,64
    conv1 = BatchNormalization()(
        Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(inputs))
    conv1 = BatchNormalization()(
        Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv1))
    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
    #第二层 128,128
    conv2 = BatchNormalization()(
        Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(pool1))
    conv2 = BatchNormalization()(
        Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv2))
    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
    #第三层 256,256,256
    conv3 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(pool2))
    conv3 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv3))
    conv3 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv3))
    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
    #第四层 512,512,512
    conv4 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(pool3))
    conv4 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv4))
    conv4 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv4))
    pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
    #第五层 512,512,512
    conv5 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(pool4))
    conv5 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv5))
    conv5 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv5))
    pool5 = MaxPooling2D(pool_size=(2, 2))(conv5)
    #decode
    #上采样
    up1 = UpSampling2D(size=(2, 2))(pool5)
    #第六层 512,512,512
    conv6 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(up1))
    conv6 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv6))
    conv6 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv6))
    up2 = UpSampling2D(size=(2, 2))(conv6)
    #第七层 512,512,512
    conv7 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(up2))
    conv7 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv7))
    conv7 = BatchNormalization()(
        Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv7))
    up3 = UpSampling2D(size=(2, 2))(conv7)
    #第八层 256,256,256
    conv8 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(up3))
    conv8 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv8))
    conv8 = BatchNormalization()(
        Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv8))
    up4 = UpSampling2D(size=(2, 2))(conv8)
    # 第八层 256,256,256
    conv9 = BatchNormalization()(
        Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(up4))
    conv9 = BatchNormalization()(
        Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9))
    up5 = UpSampling2D(size=(2, 2))(conv9)
    #第九层 64,64
    conv10 = BatchNormalization()(
        Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(up5))
    conv10 = BatchNormalization()(
        Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv10))

    # softmax输出层
    conv11 = Conv2D(1, 1, padding='same', activation='sigmoid')(conv10)

    model = Model(inputs=inputs, outputs=conv11)

    return model