To solve the problem of traditional beacon laser spot detection algorithms being susceptible to complex background interference during the initial capture stage of satellite laser communication. YOLOv5s neural network is used to optimize and improve the initial pointing scene of satellite platforms. Selecting the original loss function with the smoothed intersection over union (SIoU) loss function and replacing the original upsampling structure with a lightweight content aware feature recombination (CARAFE) upsampling structure, adding convolutional block attention module (CBAM) attention mechanism to C3 layer, using SimSPPF to replace the original structure, and adding Coordconv structure that is conducive to perceiving position information. The improved neural network has better accuracy than traditional coarse tracking beacon spot detection algorithms, and can accurately detect the position of the spot in complex backgrounds. It is suitable for beacon spot detection in the initial capture stage and coarse tracking stage. The optimized YOLOv5s neural network achieves a precision rate of 99.7%, a recall rate of 99.3%, and exceeds the average accuracy (mAP) @0.5 by 99.7% and mAP@0.5∶0.95 by 74%.