Performance Evaluation of HDR Image Reconstruction Techniques on Light Field Images

The reconstruction of high dynamic range (HDR) images from low dynamic range (LDR) images is a challenging task. Multiple algorithms are implemented to perform the reconstruction process for HDR images and videos. These techniques include, but are not limited to reverse tone mapping, computational photography and convolutional neural networks (CNNs). From the aforementioned techniques, CNNs have proven to be the most efficient when tested against conventional 2D images and videos. However, at the time of this paper, applying such CNNs to light field contents have not yet been performed. Light field images impose more challenges and difficulties to the proposed CNNs, as there are multiple images for the creation of a single light field scene. In this paper, we test some of the existing HDR CNNs (ExpandNet, HDR-DeepCNN and DeepHDRVideo) on the Teddy light field image dataset and evaluate their performance using PSNR, SSIM and HDR-VDP 2.2.1. Our work addresses both image and video reconstruction techniques in the context of light field imaging. The results indicate that further modifications to the state-of-the-art reconstruction techniques are required to efficiently accommodate the spatial coherence in light field images.