Improved DeepLabV3+ Network Beacon Spot Capture Methods

In long-range laser communication, adaptive optics tracking systems are often used to achieve high-precision tracking. When recognizing beacon spots for tracking, the traditional threshold segmentation method is highly susceptible to segmentation errors in the face of interference. In this study, an improved DeepLabV3+ network is designed for fast and accurate capture of beacon spots in complex situations. In order to speed up the inference process, the backbone of the model was rewritten as MobileNetV2. This study improves the ASPP (Atrous Spatial Pyramid Pooling) module by splicing and fusing the outputs and inputs of its different layers. Meanwhile, the original convolution in the module is rewritten as a depthwise separable convolution with a dilation rate to reduce the computational burden. CBAM (Convolutional Block Attention Module) is applied, and the focus loss function is introduced during training. The network yields an accuracy of 98.76% mean intersection over union on self-constructed beacon spot dataset, and the segmentation consumes only 12 milliseconds, which realizes the fast and high-precision capturing of beacon spots.