Figure 4

The structure of SE-ResNet-34 network. It comprised one convolutional layer, SE-ResNet Module, a global average pooling layer (GAP), and a fully-connected (FC) layer. In the SE-ResNet block, input features went through two convolutional layers and were added with original input features. All convolutional layers were followed by batch normalization (BN) layer and ReLU. Then, the features were fed into the GAP layer, and went through two FC layers followed by ReLU and sigmoid activation function to generate per-channel weights. The output of SE block is obtained by multiplying input features with the learned weights.