Joint Mask–Dictionary Optimization for Computational Light Field Imaging

Compressive light field (CLF) imaging relies on the interplay between sensing and sparse representation, where the coherence between the measurement matrix and the reconstruction dictionary fundamentally determines reconstruction performance. Existing approaches generally optimize either the sensing mask or the reconstruction dictionary separately, leading to limited compatibility between acquisition and reconstruction. To address this issue, this paper proposes a mutual coherence-driven joint optimization framework for computational light field imaging. Specifically, an optimized sensing mask is first designed by minimizing its mutual coherence with a predefined dictionary, producing a measurement matrix better suited for compressive acquisition. A dictionary optimization model is then developed for fixed-mask imaging by jointly considering reconstruction fidelity and measurement consistency under a mutual coherence constraint. The resulting alternating optimization framework effectively improves the compatibility between the sensing process and sparse representation while preserving the practical imaging configuration. Unlike generic sensing-matrix design, the proposed formulations explicitly account for the repeated and spatially shifted mask structure induced by CLF image formation. Experiments on five Stanford light field scenes at sampling ratios of 1/4 and 1/9 show that the optimized mask improves the mean PSNR over the uniform-random mask by 1.21 dB and 2.21 dB, respectively, while the optimized dictionary improves the mean PSNR over K-SVD by 0.70 dB and 0.83 dB, respectively. These results demonstrate the effectiveness of adapting either the coded mask or the reconstruction dictionary to the CLF measurement model.