Global attention network with multiscale feature fusion for infrared small target detection

Abstract

A global attention network (GANet) with multiscale feature fusion is proposed to detect infrared small target by introducing a transformer attention module and an adaptive asymmetric fusion module. The transformer attention module is designed to learn the long-range relationship between small targets and background. The adaptive asymmetric fusion module is employed to aggregate the multiscale contextual information from high-level and low-level features. In addition, a target duplicating data augmentation strategy by copy-pasting small targets many times is proposed to increase the positive samples during training for suppressing the class-imbalance problem. Extensive experiments on infrared small target datasets demonstrate that our method can achieve high detection accuracy and low false alarm rate compared with some state-of-the-art model-driven and data-driven methods.