Cross-region unsupervised denoising for high fidelity integral imaging 3D display

Restoring high-fidelity light fields from noisy element image arrays (EIAs) is essential for practical integral imaging (InIm) 3D display. Existing methods mainly rely on correlations between locally adjacent views, which limit their ability to capture long-range geometric consistency and make them vulnerable to complex noise distributions. This work presents CRN2N, a physics-aware unsupervised denoising framework that exploits cross-region correlations across the EIA. Starting from a physics-based formulation of the mixed Poisson-Gaussian degradation encountered in the imaging acquisition chain, the proposed scheme constructs unsupervised training signals by mining statistical relationships among cross-region observations, thereby eliminating the need for clean ground truth or explicit noise-parameter estimation. We further develop a geometry-aware attention residual denoising network to model long-range spatial and angular dependencies, enabling effective noise suppression while preserving high-frequency texture details and multi-view geometric consistency. Experimental results show that CRN2N achieves competitive or superior performance under various noise conditions and 3D display scenarios. Under mixed Poisson-Gaussian noise, CRN2N improves the average PSNR by 13.14 dB and increases the average SSIM by 212.7% compared with the noisy input. These results indicate that the method provides a practical framework for high-fidelity unsupervised light field restoration and offers potential for future extension to dynamic and robust real-time InIm 3D display.