PAU2-Net: local and global interaction for infrared small target detection

Abstract

ABSTRACT Infrared small target detection (ISTD) is a significant technique for search and rescue applications. The lack of intrinsic target features and complex backgrounds make small target detection challenging. Furthermore, existing deep learning-based methods overlook the imbalance among categories when designing supervision method. For robust detection, we propose a Pyramid Attention U-shaped network (PAU2-Net) in this paper. Specifically, we designed a pyramid attention encoder to establish long-range channel dependencies and enhance the interaction between local and global information. It extracts small target features and enhances high-level understanding of the scene at the same time, which helps reduce false alarm rates. In our proposed adaptive multiscale supervision method, we generate a group of multiscale labels by pooling to guide the deep outputs of the network. The labels maintain the target position information at different scales and help preserving the target information as the network deepens. Comparison experiments on SIRST-V2 and NUDT-SIRST datasets show that our method can achieve superior performance in terms of detection rate and precision compared with the state-of-the-art algorithms.