SR-HDRNet: A Unified Network for Single-Image Super-Resolution and HDR Enhancement

Abstract

Super-resolution (SR) and high-dynamic-range (HDR) enhancement are different image restoration tasks and require specialized models to be trained for each task. This work aims to tackle the all-in-one enhancement task that handle both super-resolution and HDR enhancement with a single model. To facilitate research, we construct a new dataset, namely SR-HDRSet, containing over 5,000 training samples. Inspired by the conventional 3D lookup table (3D-LUT) theory, deep network, and guided filter, we further propose a unified network for joint super-resolution and HDR image restoration model called SR-HDRNet. First, we integrated neural networks with 3D-LUT to create an image-guided 3D-LUT learnable module that expanded the color gamut and dynamic range. Next, we introduced a novel deep multi-scale feature improvement module designed to extract multi-scale texture features and efficiently expand the receptive field. Finally, we developed a learning-guided filter (LGF) module that enhanced the edge details, leading to the final reconstructed HR-HDR image. On the 2× and 3× experiments, the mean PSNR values of the proposed model surpassed the other state-of-the-arts methods by 0.61-1.43 dB and by 0.26-1.71 dB, respectively. Meanwhile, the SSIM also achieved competitive values. The comparisons validate that our proposed framework outperforms the current state of the art.