Dual-IRDet: Cross-attention-based dual-band infrared images fusion for aircraft anti-interference detection

This paper establishes anti-interference detection models for both single-band and fused dual-band infrared images to address the anti-interference problem where aerial infrared targets employ continuous infrared decoy deployment to disrupt the infrared detector's locking, tracking, and to mislead aircraft away from the target. An anti-interference detection model is established based on cross-feature fusion of dual-band infrared features. Firstly, a dual-branch backbone network is designed to extract features from dual-band infrared images, which can independently extract feature information from each band. Secondly, a segment-transform-fuse feature extraction strategy is developed to remove the redundant information in the output feature maps from single-channel infrared images after convolutional layers, which enhances feature representation by constructing inter-channel correlations while reducing redundancy in feature channels. The backbone network reuses the feature extraction strategy multiple times, thus establishing a more efficient and streamlined model. Finally, to capture more complementary information between the two infrared bands, a cross-fusion module is designed to learn the complementary relationships between mid-wave and long-wave infrared features, which models long-range dependencies across bands. The results on the constructed dual-band infrared simulation dataset demonstrate that the proposed target anti-interference detection model based on dual-band infrared images achieved an average anti-interference detection accuracy of 88.6 %, which enhances the identification efficiency of the single-band model YOLOv7 and the similar fusion detection model UA-CMDet by 3.9 % and 6.6 %, respectively.