ABSTRACT
The general U-Net architecture, first described in Ronneberger, Fischer, and Brox, has been used for countless tasks involving image segmentation, as a sub-module of numerous composite models as well as in various standalone forms. To explain the name “U-Net”, there is no better way than to reproduce a figure from the paper. The left “leg” of the U shows a sequence of steps implementing a successive decrease in spatial resolution, accompanied by an increase in number of filters. The right leg illustrates the opposite mechanism: While number of filters goes down, spatial size increases right until we reach the original resolution of the input. The decoder is made up of configurable blocks. A block receives two input tensors: one that is the result of applying the previous decoder block, and one that holds the feature map produced in the matching encoder stage.
