nekitmm/starnet

Implements a convolutional encoder-decoder residual network trained with combined L1, adversarial, and perceptual losses to remove stars from astrophotography images in a single pass. Designed for TensorFlow 2.x with pre-trained weights (~700MB) optimized for refractor telescope data, though supports fine-tuning on custom datasets to adapt for different imaging systems. Integrates into PixInsight/Photoshop workflows as a preprocessing step for nebulosity enhancement before classical star removal techniques.

306 stars. No commits in the last 6 months.