Example architectures – FCN and U-Net

Most convolutional encoders-decoders follow the same template as their fully connected counterparts, but leverage the spatial properties of their locally connected layers for higher-quality results. A typical convolutional AE is presented in one of the Jupyter notebooks. In this subsection, we will cover two more advanced architectures derived from this basic template. Both released in 2015, the FCN and U-Net models are still popular, and are commonly used as components for more complex systems (in semantic segmentation, domain adaptation, and others).