An Infrared Target Detection Method Based on Dual Attention Network for Feature Distillation

Due to the lack of infrared image data, infrared target detection is still a challenging task. To increase the performance of infrared target detection, this paper proposes a feature distillation method based on dual attention meta-network. Specifically, a dual attention meta-network is designed to transfer feature knowledge from visible to infrared images. The network includes two parts: Spatial Attention Meta-network (SAM) and Channel Attention Meta-network (CAM). In SAM, the distance between visible and infrared features is minimized by designing L2 loss, so that the knowledge of visible features contained in the teacher model is transferred to the student model. In CAM, the similarity between visible and infrared features is calculated as the distillation strength of feature connections to achieve more efficient feature distillation. In addition, to avoid the transfer of negative information, a feature selection method based on mutual information evaluation is proposed, in which the feature connections with larger mutual information are retained by setting a mutual information threshold. Experiments on public infrared detection dataset show that the proposed method outperforms the state-of-the-art feature distillation methods. In particular, the proposed method achieves an accuracy of 94.35% in infrared target detection, which is a 5.85% improvement compared to one without feature distillation.