Abstract:This paper studies an anti-interference recognition algorithm based on infrared dual-band feature cross-fusion to address the problem of infrared aerial targets being disturbed by infrared decoys. First, a dual-branch backbone network is designed to extract features from mid-wave infrared (MWIR) and long-wave infrared (LWIR) images separately. To reduce redundant information in the convolutional layer output of single-band images, a segmentation-transformation-fusion feature extraction strategy is proposed. This strategy improves feature representation, reduces channel redundancy, and is reused multiple times in the backbone network to enhance efficiency and compactness. Second, a cross-fusion module is constructed to explore complementary information between the two infrared bands. This module models the long-range dependency of cross-band features and improves resistance to infrared decoy interference. It effectively captures the complementary relationship between MWIR and LWIR features, enhancing target recognition stability. Finally, the experimental results on a simulated infrared dual-band image dataset show that the proposed algorithm achieves an average anti-interference recognition accuracy of 81.8%, which is 3.3% higher than YOLOv7.