HSTNet: A Hybrid Spatial–Channel Sparse Transformer Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) has many applications in multiple secure fields, e.g., anti-Unmanned Aerial Vehicle systems, and low-altitude threat perception. Previous works improve the performance of IRSTD, mainly focusing on global information modeling and multi-scale feature fusion. However, these works inevitably ignore the distinction between targets and backgrounds. To address this problem, we propose an effective Hybrid Spatial-Channel Sparse Transformer network, HSTNet. Specifically, we first propose a hybrid spatial-channel sparse transformer (SCST) module to sparsely model the relationship between targets and background, effectively maintain long-range dependencies. Secondly, to preserve small target details during the feature compression process, we introduce a multi-scale detail enhancement (MSDE) module. Thirdly, we propose a scale-location aware joint (SLJ) Loss to improve target perception at various scales and locations. Furthermore, to enhance the diversity and quantity of the dataset, we develop the IRSTD-Large dataset, comprising 19,558 annotated infrared images with diverse backgrounds. Finally, extensive experiments and comparisons are conducted on multiple dominant IRSTD datasets, e.g., NUAA-SIRST, IRSTD-1k, and IRSTD-Large. The results show that the proposed network surpasses current promising methods and achieves the SOTA performance. The code is available at https://github.com/juranccc/HSTNet.