Real-time optical 3D reconstruction based on Monte Carlo integration and recurrent CNNs denoising with the 3D light field display.

Abstract

A general integral imaging generation method based on the path-traced Monte Carlo (MC) method and recurrent convolutional neural networks denoising is presented. According to the optical layer structure of the three-dimensional (3D) light field display, screen pixels are encoded to specific viewpoints, then the directional rays are cast from viewpoints to screen pixels to preform the path integral. In the process of the integral, advanced illumination is used for high-quality elemental image array (EIA) generation. Recurrent convolutional neural networks are implemented as an auxiliary post-processing for the EIA to eliminate the noise of the 3D image in MC integration. 4K (3840 × 2160) resolution, 2 sample/pixel and the ray path tracing method are realized in the experiment. Experimental results demonstrate that the structural similarity metric (SSIM) value and peak signal-to-noise ratio (PSNR) gain of the reconstructed 3D image and target 3D image exceed 90% and 10 dB within 10 frames, respectively. Besides, real-time frame rate is more than 30 fps, showing the super efficiency and quality in optical 3D reconstruction.