HybrMamba: A Frequency-Aware Spatial Collaborative Network With Gated Physical Prior Head for Infrared Image Denoising

Abstract

Infrared images are frequently degraded by structured stripe noise, blind pixels, and complex background interference caused by detector nonuniformity and circuit imperfections, making it challenging to simultaneously achieve global artifact suppression and local detail preservation. To address this issue, we propose a frequency-aware hybrid collaborative network for infrared image denoising, termed HybrMamba, which combines physical degradation priors with efficient neural modeling. The proposed architecture is fundamentally constructed by two macro-components: a gated physical prior head (GPH) at the front end, and a core backbone composed of multiple cascaded HybrMamba blocks. The GPH leverages median-filtering statistics to adaptively correct blind pixels at the input stage. Each HybrMamba block adopts a tri-branch architecture, including a global structural component (GSC) branch for long-range modeling, a local convolutional extractor (LCE) branch for fine-grained feature extraction, and a spatial–frequency gating (SFG) branch. The SFG employs 2-D fast Fourier transform (FFT) to identify stripe-related spectral components and generate adaptive gating masks for feature modulation. A cross-channel feature alignment (CFA) module is further introduced to fuse multidomain features. Experiments on public benchmarks show that HybrMamba achieves 39.25-dB peak signal-to-noise ratio (PSNR) on CVC-14 and 36.04 dB on FLIR v2. Zero-shot evaluation on a real-world dataset of 200 no-reference infrared images further demonstrates its robustness and generalization capability in practical scenarios.