Among the solutions inspired by FCNs, the U-Net architecture is not only one of the first; it is probably the most popular (proposed by Olaf Ronneberger, Philipp Fischer, and Thomas Brox in a paper entitled U-Net: Convolutional networks for biomedical image segmentation, published by Springer).
Also developed for semantic segmentation (applied to medical imaging), it shares multiple properties with FCNs. It is also composed of a multi-block contractive encoder that increases the features' depth while reducing their spatial dimensions, and of an expansive decoder that recovers the image resolution. Moreover, like in FCNs, skip connections are used to connect encoding blocks to their decoding counterparts. The decoding blocks are thus provided with both the contextual information from the preceding block and the location information from the encoding path.
U-Net also differs from FCN in two main ways. Unlike FCN-8s, U-Net is symmetrical, going back to the traditional...
