Lightweight Zero-Shot DoLP Denoising via Adaptive Masking and Directional Convolution

Infrared polarimetric imaging offers advantages over conventional intensity-based modalities by revealing surface roughness and material properties. In this context, the Degree of Linear Polarization (DoLP), derived from polarization-resolved intensity measurements, is highly sensitive to surface geometry and material differences. As targets and backgrounds often exhibit distinct polarimetric responses under identical illumination, DoLP effectively amplifies these differences, thereby enhancing target-background contrast. In practical imaging systems, DoLP is derived through multi-channel polarization measurements and subsequent computational reconstruction. However, DoLP is inevitably degraded by hybrid noise, including signal-dependent shot noise and pronounced spatially correlated stripe noise, which severely degrade image quality and undermine the stability of downstream perception tasks. These complex noise patterns render data-hungry supervised models ineffective due to the lack of clean ground truth, while standard zero-shot methods often fail because strong spatial correlations violate their assumption of pixel-wise independence. To address these challenges, we propose a lightweight zero-shot self-supervised framework for DoLP denoising, termed Lightweight Adaptive Masking Directional Network (LAMD-Net). The proposed method employs a local entropy-guided adaptive masking strategy and integrates multi-scale directional convolutions within a parameter-efficient TinyUNet backbone to suppress stripe noise with low computational overhead. Extensive experiments on a custom long-wave infrared polarimetric dataset demonstrate that LAMD-Net outperforms state-of-the-art zero-shot baselines in both visual quality and efficiency.