Abstract
In a study of image quality using local, adaptive, contrast enhancement, we noted and investigated how responses varied between subjects and varied with video content.
Forty normally sighted subjects made pair-wise comparisons of side-by-side views of HD video enhanced at four levels (off, low, medium, and high) by two PureAV RazorVision devices, each separately connected to one of two side-by-side 42″ LCD HDTVs. We used logistic regression to derive Thurstone-like preference scales for the enhancement levels.
After each session subjects were asked to explain their preferences. Responses fell into two broad categories, with some subjects preferring enhanced video whereas others did not. This was reflected in their preference scales, which showed substantial, repeatable, individual differences. From each subject's comparisons of 64 video clips an enhancement preference score (EPS) was calculated. EPS was distributed bi-modally.
Subjects also indicated that video content was important to their enhancement preferences. Thus, in a post-hoc investigation of video content, we re-calculated EPS twice for each subject, once for videos that had primarily human faces, and second for videos with minimal face content (“non-face”). EPS differed between face and non-face videos, with less enhancement preferred for videos with substantial face content. This was true for most subjects (p=.005), and did not depend on whether subjects generally preferred enhancement or not.
mage quality measurement may be complicated by individual preference differences and by video content. This suggests that image quality metrics need to consider content and that a single model or computational metric for preferred image quality may not be representative of all viewers.
NIH grants EY05957, EY16093, and Analog Devices, Inc.