Machine learning based chroma phase offset detection and correction in motion video

Abstract

Generally, chroma phase or hue offset issues within a scene are hard to detect, without a reference or context (i.e. some apriori knowledge about how certain objects within the scene should actually appear in terms of their hue). Moreover, when it comes to skin/flesh tones, hue deviation can be noticeable and can markedly degrade the viewer quality of experience(QoE), whenever it does occur. However a lot of research has gone into flesh tone detection, specifically, the color gamut within which flesh tone is present. This topic has been well documented in the literature with respect to various color spaces: red, green, blue (RGB) and YIQ. Therefore, overall issues with chroma offset or hue within the video content could potentially be approached by extracting and analyzing a reliable reference, such as skin or flesh tone (if present), within some allowable deviation. This involves machine learning (ML) based facial recognition and tracking followed by skin tone region recognition within the detected facial sequence (i.e. Region of Interest). The skin region serves as a ‘self-reference’ in order to discern any inherent phase offset within the content. Finally, the angular chroma deviation discerned can then be used for subsequent correction as well.