The channel excitation module is implemented by a 1 × 1 Depth-wise convolutional layer. Because each convolution kernel acts on only one channel, it is easier to correct the weights of each channel. Finally, the channel-corrected feature map is obtained by multiplying the input feature map ...