基于混合上下文注意力网络的短波红外星点 检测算法

To address the problem of low detection accuracy for point targets and trailing targets in dim space backgrounds during star detection tasks, a mixed context attention network (MCANet) algorithm is proposed. Firstly, a lightweight feature pyramid network structure suitable for small star target detection is constructed, where effective target information is retained by designing network downsampling and feature combination strategies. Secondly, a mixed context attention mechanism (MCAM) is introduced, employing differentiated enhancement methods for high-level semantic and low-level detail features to achieve efficient utilization of target features. Finally, a soft intersection-over-union loss combined with Hausdorff distance (HD) is utilized to enhance the network’s ability to capture boundary and shape details, thereby improving overall perception of trailing targets. Experimental results on the constructed measured star dataset demonstrate that the proposed algorithm achieves optimal detection performance across all test sets with an optimal parameter quantity of 0. 26 × 10⁶.