MOU-Mamba: Multi-Order U-shape Mamba for infrared small target detection

Abstract

Infrared small target detection(IRSTD) algorithms based on CNN and Transformer have achieved remarkable results. However, approaches relying on CNNs are prone to missed or false detections due to the inherent limitations of their local receptive fields, while Transformer-based structures are insufficient in local feature extraction and are computationally expensive. Recently, different variations of the Mamba, especially the integration of the 2D-Selective-Scan (SS2D) technique, have posed strong competition to CNN– and Transformer-based methods. Therefore, this paper proposes a Multi-Order U-shape Mamba (MOU-Mamba) network to explore the potential application of SS2D in IRSTD. Specifically, we design a Multi-Order 2D-Selective-Scan (MO-SS2D) module that gradually minimizes redundant information produced by SS2D operations via multi-order interactions, effectively suppressing background interference. Moreover, we propose a Local-Guided 2D-Selective-Scan (LG-SS2D) module that enhances target discernibility by constructing complementary local and global features at each interaction level. Experimental results from three public datasets show that the proposed MOU-Mamba exceeds some state-of-the-art (SOTA) methods, maintaining an effective balance between accuracy and computational complexity.