LHDR: HDR Reconstruction for Legacy Content using a Lightweight DNN

TL;DR

This paper introduces LHDR, a lightweight deep neural network designed for HDR reconstruction from legacy SDR content, achieving high performance with minimal computational cost and addressing limitations of existing methods.

Contribution

The paper presents a novel lightweight DNN architecture specifically tailored for HDR reconstruction of degraded legacy SDR images, emphasizing degradation modeling.

Findings

Achieves competitive HDR reconstruction quality with reduced computational complexity.

Effectively handles various degradation types in legacy SDR content.

Outperforms existing methods in speed and efficiency.

Abstract

High dynamic range (HDR) image is widely-used in graphics and photography due to the rich information it contains. Recently the community has started using deep neural network (DNN) to reconstruct standard dynamic range (SDR) images into HDR. Albeit the superiority of current DNN-based methods, their application scenario is still limited: (1) heavy model impedes real-time processing, and (2) inapplicable to legacy SDR content with more degradation types. Therefore, we propose a lightweight DNN-based method trained to tackle legacy SDR. For better design, we reform the problem modeling and emphasize degradation model. Experiments show that our method reached appealing performance with minimal computational cost compared with others.