U-Net

Best for: Pixel-level segmentation

How it works

An encoder-decoder for dense segmentation. The contracting (encoder) path downsamples with convolutions to capture context, caching features $e^{(l)}$ at each scale; the symmetric expanding (decoder) path upsamples and concatenates the matching encoder features via skip connections, $u^{(l)}=\text{Conv}\bigl(\text{Concat}(u^{(l-1)},\text{crop}(e^{(l)}))\bigr)$, restoring spatial detail lost during downsampling. A final $1\times1$ convolution maps each pixel to class logits, giving accurate segmentation even from modest training sets.

Common fields

Medical imaging · satellite imagery