Antiocclusion Infrared Aerial Target Recognition with Vision-Inspired Dual-Stream Graph Network

Infrared-guided aircraft usually faces complex interference, such as natural backgrounds and infrared decoys, which decrease the performance of infrared target recognition models. A great challenge is the strong interference posed by the combination of target maneuvering and the dense and continuous deployment of infrared decoys, severely limiting the target recognition and anti-interference capabilities of infrared imaging seekers. To address the issue of target occlusion in the imaging phase, this article proposes an antiocclusion recognition model. First, inspired by the ventral pathway's capabilities of shape information processing, a skeleton-based shape feature description model is proposed based on target keypoints, which incorporates the prior knowledge about the spatial relationships of skeleton nodes inherent in the shape features of the target. Second, inspired by the dorsal pathway's capabilities of motion information processing, a motion-based feature description model is proposed using optical flow, which improves keypoint estimation performance by integrating motion information. Moreover, considering the influence of local keypoint occlusion on the stability of the target's skeleton shape representation, a learning module is introduced, which transfers high-order semantic information between nodes to strengthen the connections between keypoints while suppressing the transmission of meaningless features. Finally, an antiocclusion recognition model is proposed based on a vision-inspired dual-stream graph network (VIDSGN) for environments with interference and occlusion, which can estimate the positions of occluded nodes using known keypoints and high-order information. The experimental results on the infrared dataset constructed in this article indicate that our model has stable representation capabilities for target shape structures in the presence of partial occlusion.