关键词: 深度学习, 红外小目标检测, 双编码, 空频混合, 跨层引导

Abstract: Infrared Small Target Detection (IRSTD) aims to accurately find targets from infrared images with low signal-to-clutter ratio, and has been widely used in many fields. However, due to the weak target features and severe background interference, existing methods struggle to effectively extract the structural information of the target. This leads to issues such as incomplete target segmentation and low detection accuracy. Moreover, these models usually have a large number of parameters. To overcome above problem, a dual-coding space-frequency mixing IRSTD method was proposed. Firstly, using U-Net3+ as the basic framework, a dual-coding structure combining Multi-Shape Context Aware module and Frequency-Domain Interactive Attention module was proposed to extract space-frequency mixing features in the coding stage. Secondly, in the decoding stage, a Cross-Layer Feature Guide module is designed to fuse multi-scale feature maps. The proposed method is experimentally verified on NUAA-SIRST and IRSTD-1k datasets, the number of parameters is 0.86 M, and the Intersection over Union (IoU) values reach 78.11% and 69.08% respectively. Compared with the Attention Multiscale Feature Fusion U-Net (AMFUNet), the number of parameters is reduced by 1.31 M, and the IoU values are increased by 2.25 percentage points and 1.23 percentage points respectively. The experimental results show that the proposed method has high detection performance while retaining fewer parameters.

Key words: deep learning, infrared small target detection, dual-coding, space-frequency mixing, cross-layer guide