Window size influence on SSIM fidelity

It has been claimed that to calculate local Structural Similarity (SSIM) index, it is better to employ an 11×11 Gaussian window (called standard SSIM index). In this paper, this claim will be evaluated and shown that contrary to popular belief, not only the standard SSIM index is not always well correlated to the MOS scores but also in some distortion types, its illusory results may be deceiving and cause some wrong comparisons. To enhance the Pearson Cross Correlation (PCC) score, the SSIM indices should be tuned such that a clear and sharp relationship can be found between the SSIM indices and the MOS scores. This paper employs the local window size as a tuning tool and studies its influences on the PCC score. The simulation results show that larger local window is better for distortions with larger signal-error cross-correlation while smaller ones are better for distortions with lower cross-correlation.