Denoising 3D integral images by a single-shot unsupervised deep neural network

Integral imaging is a passive three-dimensional imaging technique that captures the radiance field through an array of apertures, offering promising applications in various fields such as 3D microscopy, metrology, and augmented reality, amongst others. However, integral imaging often suffers from noise and deterioration of image quality that can compromise the 3D image feature extraction or visualization. Current state-of-the-art denoising methods for integral imaging rely heavily on deep neural networks. Yet, these methods face a significant hurdle due to the scarcity of ground-truth integral imaging databases required for supervised learning. Furthermore, they are limited only to the imaging conditions of the predetermined training data. In this paper, we propose a new unsupervised deep learning method for integral imaging denoising using a single integral image shot. Our single-shot Noise2Noise method exploits the inherent similarity between elemental and integral imaging. Thus, it adapts to the particular image conditions during the acquisition and can increase the image quality without needing clean images or prior knowledge of the noise model.