UP-Dem: Deep Unrolling Convolutional Sparse Coding for Color Polarization Image Demosaicking

Division-of-focal-plane polarimeter (DoFP) cameras enable real-time polarization imaging by simultaneously capturing multiple polarization orientations at the cost of spatial resolution. Existing conventional and deep learning-based reconstruction methods still struggle to robustly estimate Degree of Polarization (DoP) and Angle of Polarization (AoP) under real-world conditions. Towards this end, we propose UP-Dem, a CSC-inspired deep unrolling framework integrating convolutional sparse coding (CSC) and polarization-specific constraints for color polarization demosaicking. Specifically, CSC provides model-driven reconstruction priors, while learnable modules and polarization-oriented losses improve data-driven adaptability by preserving DoP/AoP coherence and suppressing polarization artifacts. Through a CID/PID two-stage reconstruction pipeline and Stokes Feature Injection, the proposed framework progressively recovers chromatic/intensity structures and polarization-dependent information, achieving both interpretability and adaptability. Extensive experiments verify the effectiveness of our method, which achieves state-of-the-art (SOTA) results on DoTP benchmarks and real-world DoFP data. Moreover, we construct a new dataset dubbed OPID to address the lack of high-quality outdoor data in existing benchmarks and evaluate the performance of various models in outdoor environments. Our UP-Dem leads to superior AoP estimation accuracy and strong performance on downstream tasks including surface normal estimation and de-reflection. Ablation studies confirm the effectiveness of the proposed architecture and loss design. Code and dataset will be made publicly available at: https://github.com/roydon-luo/UP-Dem.