Single-Image HDR Reconstruction With Task-Specific Network Based on Channel Adaptive RDN

Guannan Chen, Lijie Zhang, Mengdi Sun, Yan Gao, Pablo Navarrete Michelini, YanHong Wu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 398-403

We describe our solution for the NTIRE--2021 High Dynamic Range Challenge with Single Frame Track where we achieved the 3rd place in terms of muPSNR and the 1st place in terms of PSNR. Aiming at this challenge we introduce the Task-specific Network based on Channel Adaptive RDN(TCRDN) that achieves good performance on the similarity with the ground truth. The network is divided into three subnets: Image Reconstruction(IR), Detail Restoration(DR) and Local Contrast Enhancement(LCE). Each subnet processes its own task, and results are fused to produce the HDR output. We carefully design these subnets so that they are properly trained for their intended purpose: detail restoration in the IR subnet and contrast enhancement in the LCE subnet. The Channel Adaptive RDN is a novel network working as the subnet backbone that combines the classic Residual Dense Network(RDN) and the Gate Channel Transformation layer. The L1 loss is used for training the network and the final model can balance the trade--off between PSNR and muPSNR for high performance in the competition's task.