Tone-mapped High Dynamic Range Image Restoration with a Mask-applied Neural Network

초록

In this paper, we propose a neural network-based image restoration method for reconstructing a single tone-mapped high dynamic range image (which has higher color-reproduction rate and detail preservation than a low dynamic range image) from a single low dynamic range image, including over- and under-exposed images. Specifically, the proposed method aims to solve the problem of restoring detail information for the clipped area, in addition to the problem of restoring color information from existing over- and under-exposed images. The proposed method uses a mask and a mask-applied neural network to distinguish between clipped and non-clipped regions. In addition, it restores the clipped region based on image inpainting, and restores the non-clipped region based on image-to-image translation. The proposed method showed higher color and detail restoration for the clipped region within a certain size, compared to conventional methods, despite the relatively small number of neural network parameters. In addition, the Fréchet inception distance score and qualitative results showed that the proposed method restores the clipped region naturally without degrading the perceptual quality, compared to other methods. © 2019 Institute of Electronics and Information Engineers. All rights reserved.

키워드