GMTNet: Global Feature Integration Mamba–Transformer Hybrid Network for Infrared Small-Target Detection

Infrared Small Target detection (IRSTD) is crucial for surveillance, rescue and early warning applications, but it still faces many challenges due to weak signal, limited texture and clutteredbackground. Existing methods based on Convolutional Neural Networks (CNN) lack global context information, while methods based on Transformers and state-space models may be too complex or too smooth, resulting in missed detections. We propose a Mamba-Transformer hybrid network called GMTNet, which is able to selectively fuse global semantics and local details. GMTNet contains three key components: (1) a directional state-space Transformer module for combining direction-aware long-range modeling with fine local cues; (2) a wavelet-based multidimensional lossless downsampling module to preserve high-frequency information during resolution reduction; (3) DualAug module for combining geometric and photometric perturbations for robustness. Experiments on IRSTD-1K, NUAA-SIRST and NUDT-SIRST show that GMTNet consistently outperforms the state of the art methods and can provide robust and efficient infrared small target detection.