A TF1 U-Net that segmented galaxies and radio noise

What it does

This is a generic U-Net implementation written for TensorFlow 1. It trains convolutional networks to perform pixel-level segmentation on arbitrary imaging data, from toy problems like spotting circles in noise to production tasks such as detecting galaxies or scrubbing radio frequency interference from telescope observations.

The interesting bit

The codebase grew out of a 2017 Astronomy & Computing paper on RFI mitigation, giving it a research pedigree rare in early deep-learning tooling. It stays domain-agnostic despite its academic roots—you can point it at any imaging task without fighting architecture-specific assumptions.

Key highlights

• Handles arbitrary segmentation problems, including toy circle detection, radio-astronomy RFI mitigation, and galaxy/star detection in wide-field imaging.
• Bundles Jupyter notebooks for a toy problem and the original radio-data use case.
• Backed by a peer-reviewed paper with a ready-made BibTeX entry.
• Documentation hosted on ReadTheDocs.

Caveats

• Explicitly discontinued; the author recommends the TensorFlow 2 reimplementation at jakeret/unet.
• Built for TensorFlow 1, so it inherits that version’s end-of-life status.

Verdict

Useful if you are maintaining legacy TF1 code or reproducing the original 2017 RFI paper. For new projects, follow the author’s own advice and use the modern successor.