Single-Shot High Dynamic Range Imaging via Deep Convolutional  Neural Network

Abstract

We propose a single-shot high dynamic range (HDR) imaging algorithm with row-wise varying exposures in a single image using a deep convolutional neural network (CNN). We first convert an input raw Bayer image into irradiance values by calibrating rows with different exposures. Then, we develop a new CNN model to restore missing information resulting from under- or over-exposed pixels and reconstruct the raw radiance map. Finally, we obtain the HDR image by applying a demosaicing algorithm to the raw radiance map. Experimental results on simulated images demonstrate that the proposed algorithm provides higher quality HDR images, with more details and less artifacts, than conventional algorithms.