Simplified UNet architecture (Ronneberger et al. 2015). The complete architecture is composed of 5 UNet blocks (here only three are represented). A UNet block is composed of two convolutions 3 × 3 and Relu activation followed by a max pooling layer for the encoding part and one up-convolution layer followed by two convolutions 3 × 3 and Relu activation for the decoding part. Skip connections correspond to the concatenation operation and are represented by dashed lines. With our implementation (five blocks), we obtain an output of size 2 × 2 × 1024 after the bottleneck given an image of size 32 × 32 × 1. The input image is represented on the left, while the output is shown on the right.