As imaging becomes increasing pervasive in our daily lives, understanding image quality is more important than ever. The final mediator of image quality is the human viewer. Yet, surprisingly, the image quality measure that is currently most widely used SSIM, is not explicitly based on any aspects of the human visual system, although it is possible to establish a monotonic relationship between SSIM scores and Mean Opinion Scores of human observers. In this talk, I will discuss two image quality models recently developed in our laboratory that are explicitly based on aspects of the human visual system. The first model is intended for still images, and among other aspects, accounts for the inhibition of masking by visually recognizable structure in the image. The second model is intended for video, and accounts for the color-dependence of artifact visibility in compressed video sequences, as mediated by spatial activity and saliency. For both models, we present psychophysical data to validate the advantages of our methods, as judged by human observers.
Presented at the 11th Bay Area Multimedia Forum (BAMMF), 13 December 2016 (bammf.org)