Improving HDR Rendering with UMTMNet’s Multilayer Perception

High dynamic range (HDR) images have recently become increasingly popular because they can show details in both dark and bright regions, while mainstream displays only support low dynamic range (LDR) images and thus require tone mapping algorithms for compression. Many tone mapping operators (TMOs) based on deep learning methods have been proposed recently in the field of HDR tone mapping, but their use in HDR tone mapping is somewhat limited due to the lack of sufficient HDR-LDR training pairs. In this paper, we propose a unified framework for unsupervised learning of tone mapping based on multilayer perception (UMTMNet). In order to improve the unsupervised training, this paper improves the original DNN model by proposing a layer-to-layer based structural loss, and a multilayer perceptual loss. In addition, in order to extract the multi-scale features of the image and suppress the noise, this paper integrates the traditional pyramid model as the front-end of the network for optimization, and for the common artifacts in the tone mapping, this paper adopts the UMTMNet structure to make the corresponding improvement, and adds the attention module and the residual dense block (RRDB) to identify the artifacts and optimize the network structure. The experimental results show that the algorithm proposed in this paper is able to generate natural artifact-free mapping images and achieve good results on multiple datasets.