Multi-Level Convolutional Network for Ground-Based Star Image Enhancement

The monitoring of space debris is important for spacecraft such as satellites operating in orbit, but the background star images taken by ground-based telescopes relatively complex, including stray light caused diffuse reflections from celestial bodies Earth or Moon, interference clouds atmosphere, etc. This has a serious impact on dim and small targets. In order to solve problem posed complex background, improve signal-to-noise ratio between target this paper, we propose novel image enhancement algorithm, MBS-Net, based suppression. Specifically, network contains three parts, namely information estimation stage, multi-level U-Net cascade module, recursive feature fusion stage. addition, new multi-scale convolutional block, which can laterally fuse perceptual field information, fewer parameters fitting capability compared ordinary convolution. For training, combine simulation real data, use obtained data pre-training way parameter migration. Experiments show that algorithm proposed paper achieves competitive performance all evaluation metrics multiple datasets.