Infrared small target detection via contrast-enhanced dual-branch network

Abstract

Infrared small target detection faces challenges including small target size, low contrast, and random image distribution. To address these, this paper presents an innovative approach named Dual-Branch Contrast-Enhanced U-Net (DBCE U-Net). Building on the basic U-Net architecture, DBCE U-Net introduces a dual-branch structure and integrates a novel Contrast Conv Module. The Contrast Conv Module enhances contrast and feature representation via adaptive feature segmentation and convolution operations, while the dual-branch design combines ResNeSt modules for deep feature extraction and feature enhancement modules (RE) for primary visual perception augmentation. In comparison with other methods, DBCE U-Net leverages gradient information to enhance small target features and improves model robustness through a dual-branch structure. Experimental results demonstrate that DBCE U-Net delivers superior detection performance across challenging datasets, particularly on the NUDT-SIRST dataset, with an IoU and Pd of 94.60% and 99.05%, respectively.