Deep Learning Turns SDR Video into HDR: ACM Multimedia 2022 Breakthrough

Background

Standard‑dynamic‑range (SDR) videos have limited brightness and color gamut, while high‑dynamic‑range (HDR) videos offer a wider luminance range and richer colors, providing more realistic scene reproduction.

Motivation

Although HDR‑capable displays are becoming common, most existing video content remains in SDR format, making high‑quality HDR content scarce. Converting SDR to HDR can significantly improve user viewing experience.

Proposed SDR‑to‑HDR Algorithm

The Kuaishou audio‑video team combined deep learning techniques to develop a reverse tone‑mapping conversion algorithm. Previous methods relied on global feature modulation, which could only scale and shift features within the original SDR feature space, failing to map them to the HDR space.

To address this, the paper introduces hierarchical dynamic context feature mapping (HDCFM) and a patch‑discriminator generator (PDCG). A dynamic feature transformation module better converts SDR features to HDR features, while a hierarchical feature modulation module performs local adaptive processing for simultaneous dark‑region enhancement and bright‑region suppression.

Results

On benchmark video sequences, the proposed method outperforms the previous state‑of‑the‑art HDRTVNET (ICCV 2021) with a PSNR gain of 0.81 dB and improvements in color fidelity, visual difference, and structural similarity.

Subjective comparisons also show a noticeable quality boost, especially in color accuracy and highlight rendering, as illustrated in the following figures.