Lensless imaging provides a compact solution for visual sensing, but its reconstruction remains an inherently ill-posed inverse problem. While coded illumination has been proven to alleviate aliasing and mitigate the ill-posedness, existing approaches are confined to pixel-domain linear reconstruction, lacking deep exploration of the latent structural correlations among multiple measurements and failing to fully exploit their complementary information. This work presents a framework that integrates multi-pattern complementary coded illumination with deep feature-domain exploration. Guided by the frequency-domain correlations of the coded scenes, we design a feature decoupling and fusion network that disentangles distinct features for parallel extraction and fusion, ensuring high-fidelity recovery. Experiments demonstrate that our framework achieves superior structural and textural preservation with reduced measurement overhead, offering an efficient and robust solution for lensless imaging. • A synergistic framework integrating coded illumination with deep feature fusion. • Efficient design with 4 complementary patterns significantly reduces complexity. • Feature decoupling and parallel fusion ensure dual fidelity in structure and texture. • Validated on simulated and real data with robust performance.
Zeng et al. (Tue,) studied this question.