A Mamba-like spatio-temporal attention network for detecting infrared moving small targets

Due to the extremely limited appearance characteristics and interference from complex backgrounds, the detection of infrared small targets remains a challenge. Single-frame methods rely solely on appearance information, which is insufficient and limits performance in complex background scenes. In contrast, multi-frame methods leverage motion information in the temporal domain simultaneously and have become the focus of infrared small-target detection. Currently, multi-frame algorithms are typically based on convolutional neural network (CNN) or Vision Transformer (ViT). CNN-based methods suffer from a limited local receptive field, while ViT-based methods exhibit high computational complexity. This paper proposed a multi-frame method based on Mamba-like spatio-temporal attention network. By replacing the recurrently computed forget gate in the Mamba architecture with linear attention, the model retains Mamba’s efficient computational capabilities while becoming better suited to non-auto-regressive vision models. This paper utilizes the improved Mamba model to achieve long-range interaction and efficiently fuse spatiotemporal information across sequential images. Moreover, we further introduced an inter-frame self-attention mechanism to extract motion features from sequential images, compensating for the insufficiency of appearance information for small targets. Finally, cross-layer connections are employed to prevent loss of small target features in deep layers caused by pooling operations. Comparative experiments with state-of-the-art algorithms on two typical datasets demonstrate that the proposed algorithm exhibits significant advantages in both effectiveness and computational efficiency.