A two-stage tone mapping network based on attention mechanism for high dynamic range images

Abstract

High dynamic range (HDR) imaging enhances visual realism by capturing a wide luminance range, but displaying HDR images on devices with limited dynamic range requires effective tone mapping operators (TMOs). We propose a novel two-stage tone mapping network (TSTMNet) based on attention mechanisms to address this challenge. The first stage utilizes adaptive luminance modulation blocks based on channel attention to dynamically adjust global luminance, achieving adaptive luminance modification. The second stage integrates local enhancement transformer blocks, leveraging self-attention from Transformers to enhance local details. This combination allows the TSTMNet to utilize the strengths of both CNNs and Transformers, overcoming their individual limitations in modeling long-range dependencies and preserving local details. Extensive experiments demonstrate that TSTMNet achieves competitive performance among recent methods in both quantitative metrics and qualitative visual quality, achieving superior dynamic range compression and detail preservation. Our method offers a robust and efficient solution to the tone mapping problem. The code is available to be downloaded at https://github.com/mtlaa/TSTMNet.