High Quality Reference Feature for Two Stage Bracketing Image Restoration and Enhancement

Xiaoxia Xing, Hyunhee Park, Fan Wang, Ying Zhang, Sejun Song, Changho Kim, Xiangyu Kong; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6267-6276

In a low-light environment it is difficult to obtain high-quality or high-resolution images with sharp details and high dynamic range (HDR) without noise or blur. To solve this problem the Bracketing Image Restoration and Enhancement integrates Dnoise Deblur HDR Reconstruction and Super Resolution techniques into a unified framework. However we find that most methods select the image that aligns with GT as the reference image. Since the details of the reference image are not good enough seriously affects the feature fusion which finally leads to details being blurred. To generate a high dynamic range and a high-quality image we propose a two-stage Bracketing method named RT-IRE. In the first stage we generate the high-quality reference feature to guide feature fusion remove the degradation and reconstruct HDR to get coarse results. The second stage learns the residuals between the coarse result and the GT which further enhances and generates details. Extensive experiments show the effectiveness of the proposed module. In particular RT-IRE won two champions in the NTIRE 2024 Bracketing Image Restoration and Enhancement Challenge.