Kim Jung Hee ; Lee Siyeong ; Kang Suk-Ju

Abstract

Recently, high dynamic range (HDR) image reconstruction based on the multipleexposure stack from a given single exposure utilizes a deep learning frameworkto generate high-quality HDR images. These conventional networks focus on theexposure transfer task to reconstruct the multi-exposure stack. Therefore, theyoften fail to fuse the multi-exposure stack into a perceptually pleasant HDRimage as the inversion artifacts occur. We tackle the problem in stackreconstruction-based methods by proposing a novel framework with a fullydifferentiable high dynamic range imaging (HDRI) process. By explicitly usingthe loss, which compares the network's output with the ground truth HDR image,our framework enables a neural network that generates the multiple exposurestack for HDRI to train stably. In other words, our differentiable HDRsynthesis layer helps the deep neural network to train to create multi-exposurestacks while reflecting the precise correlations between multi-exposure imagesin the HDRI process. In addition, our network uses the image decomposition andthe recursive process to facilitate the exposure transfer task and toadaptively respond to recursion frequency. The experimental results show thatthe proposed network outperforms the state-of-the-art quantitative andqualitative results in terms of both the exposure transfer tasks and the wholeHDRI process.