UNets

A U-Net is a convolutional neural network architecture that was originally developed for biomedical image segmentation. Its distinctive feature is a symmetric encoder-decoder structure, resembling the shape of the letter "U". The encoder path, also known as the contracting path, progressively downsamples the input image through a series of convolutional and pooling layers, capturing contextual information. The decoder path, or expansive path, then progressively upsamples the feature maps while concatenating them with feature maps from the corresponding layers in the encoder path. This concatenation helps to recover fine-grained spatial details lost during downsampling, which is crucial for precise segmentation.

The skip connections, which link encoder and decoder layers, are a key innovation of the U-Net. These connections allow gradients to flow more easily during backpropagation, facilitating the learning of complex patterns. The architecture typically ends with a final convolutional layer that outputs a segmentation map, where each pixel is assigned a class label. U-Nets have proven highly effective in various image segmentation tasks beyond medical imaging, including satellite imagery analysis, object detection, and autonomous driving. Their ability to handle high-resolution inputs and produce accurate, pixel-wise predictions makes them a popular choice for many computer vision applications.