Deep Learning-Based Small Target Detection for Satellite-Ground Free Space Optical Communications

Free space optical (FSO) channels between a low earth orbit (LEO) satellite and a ground station (GS) use a highly directional optical beam that necessitates a continuous line-of-sight (LOS) connection. In this paper, we propose a deep neural network (DNN)-based small target detection method that detects the position of a LEO satellite in an infrared image, which can be used to determine the receiver alignment for establishing the LOS link. For the infrared small target detection task without excessive down-sampling, we design a target detection model using a modified ResNest-based feature extraction network (FEN), a custom feature pyramid network (FPN), and a target determination network (TDN). ResNest utilizes the feature map attention mechanism and multi-path propagation necessary for robust feature extraction of small infrared targets. The custom FPN combines multi-scale feature maps generated from the modified ResNest to obtain robust semantics across all scales. Finally, the semantically strong multi-scale feature maps are fed into the TDN to detect small infrared targets and determine their location in infrared images. Experimental results using two widely used point spread functions (PSFs) demonstrate that the proposed algorithm outperforms the conventional schemes and detects small targets with a true detection rate of 99.4% and 94.0%.