Real-time self-supervised tone curve estimation for HDR image

Abstract

High dynamic range (HDR) image acquisition technology can record real-scene information. However, most display devices only support standard dynamic range images; therefore, dynamic range compression via tone mapping operators is a key technology for HDR image visualization. Recently, deep learning has achieved significantly better results than traditional methods in tone mapping; however, some problems remain. On the one hand, the labels of datasets are usually selected from the results of existing traditional methods, which limit the quality of newly generated results. On the other hand, these algorithms require huge computational resources and cannot be practically applied. As the dynamic range can be compressed using an intuitive and simple sigmoid curve mapping, a lightweight network is designed in this study to estimate the patch-wise tone curve parameters for HDR images. Simultaneously, the differentiable approximation of tone mapped image quality assessment is introduced as a self-supervised loss term so that the method only needs HDR data as training data. Experimental results demonstrate that the proposed method achieves better results than existing methods in both objective and subjective metrics at a low computational cost.