Just a U-Net

This repository contains an implementation of a classic U-NET architecture, which is widely used for image segmentation tasks. The U-NET model is known for its ability to produce high-quality segmentation maps even with a limited amount of training data.

Reference

For more details on the U-NET architecture, please refer to the original paper: Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation [arXiv:1505.04597]

Example

After training on the HAM10000 skin injury segmentation dataset, which includes expert-drawn segmentation mask annotations, the U-Net demonstrates some capability to predict skin injury segmentations in dermatoscopic images.

Melanoma and U-Net prediction mask. Original image (left), prediction (center), superposition (right). Source for the original image ISIC Archive (Unique ID: ISIC_0000031). Original image published under Creative Commons CC-0 copyright license.

Further details

It includes Dataset classes for loading data from:

• VOC2012: A classic 1000 everyday object categories semantic segmentation dataset.
• HAM10000: A skin injury segmentation dataset.