Single Image HDR Reconstruction Using Generative Adversarial Networks

Abstract

The advances in GANs have paved the way for various methods of high dynamic range (HDR) image reconstruction. In this paper, we use the structural advantages of GAN to infer natural HDR images and reconstruct missing information from a single exposure low dynamic range(LDR) image in an end-to-end fashion, which extends the dynamic range of a given image to generate HDR image. Furthermore, we propose a novel dense feedback model and the feedback mechanism, which can make use of the high-level information to refine the shallow information in the top-down feedback stream through the global feedback and the local feedback connection. The dense connections in the forward-pass enable feature-reuse and comprehensively learn complex nonlinear relationships from LDR to HDR mapping. Experiment results demonstrate proposed method produces superior performance compared to existing state-of-the-art approaches.