Cross-Frequency Context-Guided Mamba Network for Infrared Small Target Detection

Infrared small target detection remains challenging because small targets, cloud edges, sea reflections, building heat sources, and sensor noise often share similar high-frequency responses. Existing local enhancement methods tend to amplify target-like clutter, whereas global modeling methods may dilute sparse target details. To address this issue, this paper proposes CFGMNet (Cross-Frequency Context-Guided Mamba Network), which reformulates infrared small target detection as a high-frequency candidate verification problem constrained by low-frequency background context. Specifically, the proposed CFG-Mamba module decomposes deep features into low-frequency background components and high-frequency candidate regions. Mamba is applied only to the low-frequency branch to capture long-range background dependencies, and the resulting contextual representation is used to gate high-frequency responses, thereby suppressing target-like clutter without indiscriminately enhancing local details. Furthermore, a local contrast gate, a residual attention decoder, and a decoupled prediction head are introduced to perform local saliency calibration, skip-connection noise filtering, and response verification. Experiments on NUAA-SIRST, NUDT-SIRST, and IRSTD-1K demonstrate that CFGMNet achieves a favorable balance between segmentation accuracy, false alarm suppression, and inference efficiency. In particular, CFGMNet achieves 85.27% mIoU and 93.77% F1 on NUAA-SIRST and obtains the lowest false alarm rate on IRSTD-1K, while reaching 141 FPS under forward-pass-only evaluation on an RTX 4090 GPU.