LANet: Luminance-Aware Network for SDRTV-to-HDRTV Translation
J. Y. Zhao, Hua Zhang · 2024
High dynamic range (HDR) media resources, providing more details and higher contrast compared to standard dynamic range (SDR) media resources, cannot be captured by standard digital cameras directly due to sensor limitations. However, HDR cameras are excessively expensive, and most existing media resources are still in SDR. With the development of HDR display devices, there is an intense demand to transform the existing SDR media resources into the HDR version. In this work, we propose a deep neural network, LANet, to achieve standard dynamic range television (SDRTV) to high dynamic range television (HDRTV) translation. Specifically, we propose a new luminance-aware branch with self-attention mechanisms to selectively and adaptively process different regions in SDR frames that existing methods have overlooked. Experimental results show that the proposed LANet achieves state-of-the-art performance in quantitative comparisons and the visual quality of the reconstructed HDR frames have been significantly improved.