Anti-Occlusion Infrared Aerial Target Recognition With Multisemantic Graph Skeleton Model

In photoelectric countermeasure systems, the infrared imaging of missiles is critical for automatic recognition and tracking technology of aerial targets. However, complex and newly emerging infrared interference signals severely hinder the recognition performance and lock the target ability of infrared thermal imaging systems. Although considerable progress has been achieved in the development of machine vision systems for missile detection, their performance and robustness should be improved. The brain can detect learned objects in various nonideal situations (partial occlusion and various perspectives). A novel graph network learning framework was developed for object recognition. This brain-inspired anti-interference recognition model can be used for detecting aerial targets composed of various spatial relationships. A spatially correlated skeletal graph model was used to represent the prototype using the graph convolutional network. Furthermore, a novel anti-occlusion framework based on a multisemantic skeleton graph model was proposed to overcome the discontinuity of target features caused by occlusion. In this method, the location of occluded key points was inferred by learning high-order relationships and node topology information. In this study, local image features were considered as graph nodes and a high-order relationship learning module was proposed to transfer relational information between nodes. In this module, the degree of connection between target keypoints was learned to automatically suppress the delivery of meaningless features. Second, a high-order topology learning module that simultaneously learns topological information and embeds local features was proposed to directly predict node similarity scores. Finally, extensive experiments were conducted on the constructed aerial target flight infrared dataset to validate the effectiveness of the proposed model.