【659】DeConvNet 模型说明

[1] DeconvNet 模型说明

[2] github - dongmeixu/Keras-DeconvNet-Segmentation

　　模型结构图

　　实现代码：（二分类）

from keras.models import Model
from keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, BatchNormalization
from keras.layers.convolutional import Conv2DTranspose
from keras.callbacks import ModelCheckpoint
import os
import tensorflow as tf
import pandas as pd
import numpy as np
import random
import cv2


def get_net(img_size):
    inputs = Input(img_size)
    conv1_1 = Conv2D(filters=64, kernel_size=(3, 3), padding="same", activation="relu")(inputs)
    conv1_2 = Conv2D(filters=64, kernel_size=(3, 3), padding="same", activation="relu")(conv1_1)
    conv1_2 = BatchNormalization()(conv1_2)
    pool1 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(conv1_2)

    conv2_1 = Conv2D(filters=128, kernel_size=(3, 3), padding="same", activation="relu")(pool1)
    conv2_2 = Conv2D(filters=128, kernel_size=(3, 3), padding="same", activation="relu")(conv2_1)
    conv2_2 = BatchNormalization()(conv2_2)
    pool2 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(conv2_2)

    conv3_1 = Conv2D(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(pool2)
    conv3_2 = Conv2D(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(conv3_1)
    conv3_3 = Conv2D(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(conv3_2)
    conv3_3 = BatchNormalization()(conv3_3)
    pool3 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(conv3_3)

    conv4_1 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(pool3)
    conv4_2 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(conv4_1)
    conv4_3 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(conv4_2)
    conv4_3 = BatchNormalization()(conv4_3)
    pool4 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(conv4_3)

    conv5_1 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(pool4)
    conv5_2 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(conv5_1)
    conv5_3 = Conv2D(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(conv5_2)
    conv5_3 = BatchNormalization()(conv5_3)
    pool5 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(conv5_3)

    fc6 = Conv2D(filters=4096, kernel_size=(5, 5), activation="relu")(pool5)
    fc7 = Conv2D(filters=4096, kernel_size=(1, 1), activation="relu")(fc6)

    deconv_fc6 = Conv2DTranspose(filters=512, kernel_size=(5, 5))(fc7)
    deconv_fc6 = BatchNormalization()(deconv_fc6)
    unpool5 = UpSampling2D(size=(2, 2))(deconv_fc6)

    deconv5_1 = Conv2DTranspose(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(unpool5)
    deconv5_2 = Conv2DTranspose(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(deconv5_1)
    deconv5_3 = Conv2DTranspose(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(deconv5_2)
    deconv5_3 = BatchNormalization()(deconv5_3)
    unpool4 = UpSampling2D(size=(2, 2))(deconv5_3)

    deconv4_1 = Conv2DTranspose(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(unpool4)
    deconv4_2 = Conv2DTranspose(filters=512, kernel_size=(3, 3), padding="same", activation="relu")(deconv4_1)
    deconv4_3 = Conv2DTranspose(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(deconv4_2)
    deconv4_3 = BatchNormalization()(deconv4_3)
    unpool3 = UpSampling2D(size=(2, 2))(deconv4_3)

    deconv3_1 = Conv2DTranspose(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(unpool3)
    deconv3_2 = Conv2DTranspose(filters=256, kernel_size=(3, 3), padding="same", activation="relu")(deconv3_1)
    deconv3_3 = Conv2DTranspose(filters=128, kernel_size=(3, 3), padding="same", activation="relu")(deconv3_2)
    deconv3_3 = BatchNormalization()(deconv3_3)
    unpool2 = UpSampling2D(size=(2, 2))(deconv3_3)

    deconv2_1 = Conv2DTranspose(filters=128, kernel_size=(3, 3), padding="same", activation="relu")(unpool2)
    deconv2_2 = Conv2DTranspose(filters=64, kernel_size=(3, 3), padding="same", activation="relu")(deconv2_1)
    deconv2_2 = BatchNormalization()(deconv2_2)
    unpool1 = UpSampling2D(size=(2, 2))(deconv2_2)

    deconv1_1 = Conv2DTranspose(filters=64, kernel_size=(3, 3), padding="same", activation="relu")(unpool1)
    deconv1_2 = Conv2DTranspose(filters=64, kernel_size=(3, 3), padding="same", activation="relu")(deconv1_1)
    deconv1_2 = BatchNormalization()(deconv1_2)
    output = Conv2D(1, kernel_size=(1, 1), activation='sigmoid')(deconv1_2)

    model = Model(inputs=inputs, outputs=output)
    return model


model = get_net((512,512,3))
model.summary() 

　　summary

Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_3 (InputLayer) [(None, 512, 512, 3)] 0
_________________________________________________________________
conv2d_31 (Conv2D) (None, 512, 512, 64) 1792
_________________________________________________________________
conv2d_32 (Conv2D) (None, 512, 512, 64) 36928
_________________________________________________________________
batch_normalization_22 (Batc (None, 512, 512, 64) 256
_________________________________________________________________
max_pooling2d_10 (MaxPooling (None, 256, 256, 64) 0
_________________________________________________________________
conv2d_33 (Conv2D) (None, 256, 256, 128) 73856
_________________________________________________________________
conv2d_34 (Conv2D) (None, 256, 256, 128) 147584
_________________________________________________________________
batch_normalization_23 (Batc (None, 256, 256, 128) 512
_________________________________________________________________
max_pooling2d_11 (MaxPooling (None, 128, 128, 128) 0
_________________________________________________________________
conv2d_35 (Conv2D) (None, 128, 128, 256) 295168
_________________________________________________________________
conv2d_36 (Conv2D) (None, 128, 128, 256) 590080
_________________________________________________________________
conv2d_37 (Conv2D) (None, 128, 128, 256) 590080
_________________________________________________________________
batch_normalization_24 (Batc (None, 128, 128, 256) 1024
_________________________________________________________________
max_pooling2d_12 (MaxPooling (None, 64, 64, 256) 0
_________________________________________________________________
conv2d_38 (Conv2D) (None, 64, 64, 512) 1180160
_________________________________________________________________
conv2d_39 (Conv2D) (None, 64, 64, 512) 2359808
_________________________________________________________________
conv2d_40 (Conv2D) (None, 64, 64, 512) 2359808
_________________________________________________________________
batch_normalization_25 (Batc (None, 64, 64, 512) 2048
_________________________________________________________________
max_pooling2d_13 (MaxPooling (None, 32, 32, 512) 0
_________________________________________________________________
conv2d_41 (Conv2D) (None, 32, 32, 512) 2359808
_________________________________________________________________
conv2d_42 (Conv2D) (None, 32, 32, 512) 2359808
_________________________________________________________________
conv2d_43 (Conv2D) (None, 32, 32, 512) 2359808
_________________________________________________________________
batch_normalization_26 (Batc (None, 32, 32, 512) 2048
_________________________________________________________________
max_pooling2d_14 (MaxPooling (None, 16, 16, 512) 0
_________________________________________________________________
conv2d_44 (Conv2D) (None, 12, 12, 4096) 52432896
_________________________________________________________________
conv2d_45 (Conv2D) (None, 12, 12, 4096) 16781312
_________________________________________________________________
conv2d_transpose_28 (Conv2DT (None, 16, 16, 512) 52429312
_________________________________________________________________
batch_normalization_27 (Batc (None, 16, 16, 512) 2048
_________________________________________________________________
up_sampling2d_10 (UpSampling (None, 32, 32, 512) 0
_________________________________________________________________
conv2d_transpose_29 (Conv2DT (None, 32, 32, 512) 2359808
_________________________________________________________________
conv2d_transpose_30 (Conv2DT (None, 32, 32, 512) 2359808
_________________________________________________________________
conv2d_transpose_31 (Conv2DT (None, 32, 32, 512) 2359808
_________________________________________________________________
batch_normalization_28 (Batc (None, 32, 32, 512) 2048
_________________________________________________________________
up_sampling2d_11 (UpSampling (None, 64, 64, 512) 0
_________________________________________________________________
conv2d_transpose_32 (Conv2DT (None, 64, 64, 512) 2359808
_________________________________________________________________
conv2d_transpose_33 (Conv2DT (None, 64, 64, 512) 2359808
_________________________________________________________________
conv2d_transpose_34 (Conv2DT (None, 64, 64, 256) 1179904
_________________________________________________________________
batch_normalization_29 (Batc (None, 64, 64, 256) 1024
_________________________________________________________________
up_sampling2d_12 (UpSampling (None, 128, 128, 256) 0
_________________________________________________________________
conv2d_transpose_35 (Conv2DT (None, 128, 128, 256) 590080
_________________________________________________________________
conv2d_transpose_36 (Conv2DT (None, 128, 128, 256) 590080
_________________________________________________________________
conv2d_transpose_37 (Conv2DT (None, 128, 128, 128) 295040
_________________________________________________________________
batch_normalization_30 (Batc (None, 128, 128, 128) 512
_________________________________________________________________
up_sampling2d_13 (UpSampling (None, 256, 256, 128) 0
_________________________________________________________________
conv2d_transpose_38 (Conv2DT (None, 256, 256, 128) 147584
_________________________________________________________________
conv2d_transpose_39 (Conv2DT (None, 256, 256, 64) 73792
_________________________________________________________________
batch_normalization_31 (Batc (None, 256, 256, 64) 256
_________________________________________________________________
up_sampling2d_14 (UpSampling (None, 512, 512, 64) 0
_________________________________________________________________
conv2d_transpose_40 (Conv2DT (None, 512, 512, 64) 36928
_________________________________________________________________
conv2d_transpose_41 (Conv2DT (None, 512, 512, 64) 36928
_________________________________________________________________
batch_normalization_32 (Batc (None, 512, 512, 64) 256
_________________________________________________________________
conv2d_46 (Conv2D) (None, 512, 512, 1) 65
=================================================================
Total params: 151,119,681
Trainable params: 151,113,665
Non-trainable params: 6,016
_________________________________________________________________

　　模型网络结构：