Enhancing Infrared Object Detection: An Auxiliary Multi-Head Network Approach

Abstract Infrared object detection poses unique challenges due to limited feature information compared to visible light images. This study introduces an auxiliary multi-head network enhancement method designed to improve object detection accuracy in infrared scenes. Unlike traditional approaches, our method autonomously extracts key regions of interest and employs multi-head inputs with an imbalanced network structure. By integrating a gradient-based R-IOU loss function, we address detection issues caused by object occlusion. Experimental results on public datasets demonstrate an average accuracy improvement of 2.5% across various network frameworks, highlighting the effectiveness of our proposed method.