Noise-robust holographic data storage with implicit neural representation

Holographic data storage (HDS) offers high capacity and throughput, making it a promising candidate for next-generation storage technologies. However, it is susceptible to increased bit error rate (BER) due to optical/electronic non-idealities and media inhomogeneity. To address this challenge, we propose an implicit neural representation (INR)-based data representation method with a dropout training strategy for robust HDS. In this framework, image information is encoded into compact network parameters, which are subsequently written to and read from the storage medium using an off-the-shelf HDS technique. The retrieved parameters are then loaded for inference to reconstruct the image. Compared with conventional pixel-wise image storage, the global representation enhances robustness by allowing the network to compensate for partial parameter errors. Moreover, it reduces the number of required data pages, providing compression benefits for sparse images. Both simulations and experiments demonstrate consistent gains under noisy conditions, with peak signal-to-noise ratio (PSNR) improved by over 50%, structural similarity index measure (SSIM) above 0.95, and a compression ratio of 2.92 for sparse images. These results demonstrate the effectiveness of the proposed approach for reliable HDS under high BER conditions.