SSIM Revisited

We review the design of the SSIM quality metric and offer an alternative model of SSIM computation, utilizing subband decomposition and identical distance measures in each subband. Based on this model, we also propose an extension of this metric, using a finer-grain sub-band structure resembling the spatial channel model of contrast sensitivity function (CSF). We use a weighted combination of subband-specific scores into a single error visibility score for each spatial location within an image. The averaging is calibrated based on the characteristics of the viewing setup (translation from pixels to spatial frequencies). We test this proposed model using several subjective score datasets and show that it can outperform the original SSIM and MS-SSIM in the accuracy of fit to MOS scores. We discuss applications of this proposed model for the design and improvement of video encoding applications. Specifically, we show how to design pre-filters or inloop noise shaping filters improving SSIM scores of encoded video sequences.