Lightweight visual Mamba-guided attention network for infrared small target detection

In recent years, the rapid development of basic model architectures in the field of deep learning has greatly promoted the progress of infrared small target detection (ISTD). Although convolutional neural networks (CNNs) can effectively extract local features of infrared small targets, they have shortcomings in capturing long-distance global features between the target and its background. The vision transformer (ViT) model can capture long-distance dependent and global features. Combining it with CNNs can effectively compensate for the shortcomings of CNNs in modeling long-distance global features. However, the disadvantages of large computational cost and quadratic computational complexity of the transformer model have also been inherited. Recently, the emerging Mamba model with linear complexity has attracted much attention due to its excellent ability to model long-sequence features. Many studies have shown that this model can perform well in various visual tasks with low computational burden and high accuracy. However, the direct application of Mamba alone is not effective, as it cannot fully exploit local features. We explore the effectiveness and efficiency of state-space models for single-frame ISTD tasks. We integrated the visual Mamba module into a lightweight CNN-based network for infrared small object segmentation. This integration allows the powerful ability of CNNs to model local features to be fully utilized, whereas the visual Mamba module efficiently captures long-distance global features. Therefore, our model can effectively extract both local and long-distance global feature information. Experimental results on the NUAA-SIRST and NUAA-SIRSTv2 datasets show that our model achieves better accuracy and lower computational complexity.