I was reading the U-Net paper for medical image segmentation. I had a doubt in the architecture. The authors mention that the max pooling layers in contraction path double the number feature channels while Downsampling. Can anyone explain what are feature channels and how are they doubled while max pooling? In simple language please.
The feature channels simply mean the number of channels at a given point in the network. In the U-Net architecture, the number of channels doubles after the max pooling layers, see for example the first max pooling layer. The input is of size 568 by 568 pixels with 64 (feature) channels, after max pooling this becomes an array of size 284 by 284 pixels with the same 64 channels. The next convolutional layer then keeps the same size, so 284 by 284 pixels, but doubles the number of channels to 128.