CNN-based bit error rate reduction in holographic data storage systems using simulated training datasets

In a holographic data storage system (HDSS) experimental setup, several noise sources can increase the bit error rate (BER) when we recover the data from the reconstructed data page. While some of these errors originate in the initial stages of the experimental setup and can be mitigated as the system is refined, others remain inherent to the setup itself. In this work, we present a technique based on a convolutional neural network (CNN) that enables BER testing during the early stages of the experimental setup to evaluate whether the system will meet future requirements, or even to improve the BER at a more advanced stage. For training a CNN it is crucial to have a sufficiently large dataset, for that reason we have to store and reconstruct a high volume of data pages. In the early development phases of a HDSS this process can be very time consuming because there are many steps that are not automatized. To address this challenge, we have developed a system that can simulate the recovered image that we obtain by using basic digital image processing techniques, allowing the generation of the large number of image pairs required for CNN training. After training and testing the CNN with the simulated data we have applied the method to real reconstructed experimental data pages obtained in a HDSS which incorporates, among others innovations a parallel aligned liquid crystal on silicon device spatial light modulator (PA-LCoS, SLM) as data pager, and the pages are stored in a self-made photopolymer based on polyvinyl alcohol/acrylamide (PVA/AA). The results obtained are promising and demonstrate this training approach can be successfully applied to a real-implemented HDSS.