Image quality enhancement of directional volumetric displays using generative adversarial networks

We propose a voxel optimization model utilizing generative adversarial networks for directional volumetric displays, in which the image displayed depends on the observation direction. Employing the L1 loss, learned perceptual image patch similarity (LPIPS), and adversarial losses, the proposed generator overcomes the limitation of the conventional iterative approximation method, which minimizes only pixel-wise errors. When evaluated on 40,670 image pairs from the Microsoft Common Objects in Context 2017 dataset, the generator demonstrated quantitative improvements over the conventional method, with an LPIPS of 0.1415 and a Fréchet inception distance of 5.654 (versus 0.1487 and 6.381, respectively, in the conventional method). Furthermore, visual comparisons indicated that the proposed approach mitigated the horizontal linear artifacts observed in images generated by the conventional method.