tensorflow - Keras中VDSR的剩余损失计算

Question

我正在使用带有后端张量流的 Keras。我从 ( https://github.com/GeorgeSeif/VDSR-Keras )下载了非常深的超分辨率 cnn 的代码，并尝试训练我自己的模型。该模型是用均方误差 (mse) 编译的，但原始模型有残余损失。如何在 keras 中实现残差损失。剩余损失可以计算为

loss = (label - input_image - predected_image)^2

标签： tensorflowkeras

Answer 1

VDSR 模型如下：

input_img = Input(shape=IMG_SIZE)

model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(input_img)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)

model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)

model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)

model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(1, (3, 3), padding='same', kernel_initializer='he_normal')(model)
res_img = model

output_img = add([res_img, input_img])

model = Model(input_img, output_img)

模型的输入图像是一个图像块块（一个块的分辨率为 41 x 41 像素，8 位/像素）。
在您的公式中： loss = (label - input_image - predected_image)^2 =(label - (input_image + predected_image))^2) 其中 input_image + predected_image 它是 newtork 输出。网络学习校正（line output_img = add([res_img, input_img])），以克服双三次插值的影响，这不是图像超分辨率的最佳方法。