Study of Multi-Scale temporal extension and Minkowski a MSSIM

Numerous image and video quality metrics have been proposed over the last three decades. One of such metrics is the Structural Similarity (SSIM) which deviates from the existing quality metrics that quantify image degradation on the basis of perceived errors to a new paradigm that measures image degradation based on the changes in structural information variation. A multi-scale extension to this approach measure the SSIM for the image at different scales after iteratively passing it through the low pass filter and down-sampling it. In this paper, we extend the Multi-scale Structural Similarity (MSSIM) algorithm to multiple temporal scales by using the low-pass response of the Haar wavelet and study its performance at various scales and effect of combining the individual frame quality score using Minkowski pooling at each scale. Our result in a total of 470 videos from two different publicly available database show that Minkowski´s pooling of the frames in general increases the subjective correlation, and higher temporal scales do not necessarily increase the correlation with the subjective scores.