U-Net

U-Net is a U-shaped convolutional network: it uses a symmetric encoder-decoder architecture with critical skip connections for precise pixel-level semantic segmentation.

U-Net, introduced in 2015 by Olaf Ronneberger, Philipp Fischer, and Thomas Brox, is a fully convolutional network designed for high-precision image segmentation. Its architecture is distinctively U-shaped: a contracting path (encoder) captures context, and a symmetric expansive path (decoder) enables localization. The key innovation is the extensive use of skip connections, which concatenate high-resolution feature maps from the encoder directly to the decoder. This mechanism ensures that fine-grained spatial details, often lost during downsampling, are preserved for the final segmentation mask. It was initially developed for biomedical image segmentation, proving highly effective even with limited training data.

https://arxiv.org/abs/1505.04597