Chapter 9

1. The dense blocks, where each layer is connected to all the output of the previous layers, including the input.
2. ResNet.
3. It is an adaptation of DenseNet for performing semantic segmentation.
4. Because it can be visualized like a U, with the left side downsampling and the right side upsampling.
5. No, you can just stack a series of convolutions. However, due to the high resolution, achieving a good result will be challenging, and most likely would use a lot of memory and be quite slow.
6. You can use a median blur to remove the salt and pepper noise that can be present in the segmentation masks of poorly trained networks.
7. They are used to propagate high-resolution channels and help the network achieve a good real resolution.