PFDMNet: Prior-guided frequency-decoupled Mamba network for infrared small target detection

Infrared small target detection (IRSTD) remains challenging due to the extremely low signal-to-clutter ratio and the fragile structural responses of targets. Although recent deep learning methods have achieved notable progress, most studies mainly focus on architectural modifications or feature interaction strategies, while the intrinsic physical characteristics of infrared small targets are seldom explicitly considered. Motivated by this observation, the paper proposes a prior-guided frequency-decoupled Mamba network (PFDMNet) for infrared small target detection. Specifically, a dual-domain consistency block is embedded in shallow layers, which explicitly models brightness consistency and multi-directional gradient consistency to strengthen weak target responses under cluttered backgrounds. In deep layers, the network incorporates a frequency Mamba module, which decomposes features into frequency components and applies state-space modeling to the low-frequency branch to capture long-range background structures. In addition, a role-aware fusion module is designed that treats shallow encoder features and deep decoder features differently, followed by an adaptive gating mechanism to balance shallow details and deep semantics during fusion. Extensive experiments on three public datasets, NUDT-SIRST, NUAA-SIRST, and IRSTD-1K, demonstrate that PFDMNet achieves competitive performance on multiple evaluation metrics while remaining lightweight. The relevant code will be made available at https://github.com/nuist2685/PFDMNet .