Comparison of state-of-the-art Computer Generated Holography algorithms and a machine learning approach

Abstract

This work studies the use of machine learning and, in particular, a Convolutional Neural Network (CNN) to generate digital holograms and how such a network compares to state-of-the-art iterative methods, both in terms of reconstruction quality and computation time. Since CNNs only require a single pass through the network by a target image to generate a result, and not tens or hundreds of expensive iterations as in the iterative methods, they may be able to accomplish real-time hologram generation; an ability that could open the technology to proper commercial use.In this work, a CNN built on the UNet architecture, capable of hologram generation, is presented. The network is trained on 4096 images of varying spatial frequencies, both user-generated and from the DIV2K dataset. It is compared to the most common iterative method for hologram generation, namely the Gerchberg–Saxton (GS) algorithm and its modern and improved implementations. In reconstruction quality, the neural network outperforms the original implementation of GS when evaluating Mean Square Error (MSE), geometric error (GE), Structural Similarity Index Measurement (SSIM), and Peak Signal-Noise Ratio (PSNR) of 64 unseen test images. However, on the same test images, the network lacks behind the modern, optimized GS implementations in all error and accuracy measurements. The network does, however, achieve these results at a rate 70–280 times faster than the iterative methods, depending on the particular implementation of the GS algorithm, which corresponds to a possible generation rate of the network of 32 FPS on average.