CNN based HDR/WCG Characteristic Detection for Ultra-High Definition Video

Abstract

Featuring high bandwidth and low latency, the fifth-generation (5G) network can handle Ultra-High Definition (UHD) video services including 4K and 8K with High Dynamic Range (HDR) and Wide Color Gamut (WCG). In order to ensure high quality for consumers, the parameters of UHD video contents such as resolution, bit depth, frame rate, color gamut and etc. are usually signaled during the whole workflow to keep quality consistency and interoperability. When the signaling is unavailable or unreliable, the end users are likely to obtain unpleasant quality. In this paper, we propose a novel framework for CNN based luminance and color characteristic detection to detect the actual UHD video parameters directly from pixels rather than signaled parameters, which consists of two image classification networks to capture high-level video feature representation. By utilizing Multi-Class Image Reference (MCIR) and Local Feature Extraction and Fusing (LFEF) sub-networks, our scheme performs well on color gamut and dynamic range detection. Experiments on practical video contents show that our system can achieve accuracy of 99.4% and 94.9% on color gamut and dynamic range classification separately.