# Estimating SSIM from MSE for DCT-Based Compressed Images via Modeling Local Error Statistics At [QoMEX 2026](https://qomex2026.ieee-qomex.org/), I presented the paper **"Estimating SSIM from MSE for DCT-Based Compressed Images via Modeling Local Error Statistics"** with [Maria G. Martini](https://www.kingston.ac.uk/about/staff/professor-maria-martini). Since the work was presented as a poster, I decided to go over the main ideas and results in this short video. <iframe width="560" height="315" src="https://www.youtube.com/embed/nK-LybpQMmg?si=M2oRBt4eeVqqLKsL" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe> ## Overview SSIM is one of the most widely used perceptual quality metrics for images and video. However, computing SSIM requires analyzing local statistics across the entire image, making it expensive when evaluating many candidate encodes of the same source. This talk presents **ApproxSSIMate**, an approach to estimate SSIM using only frame-level MSE while moving the expensive analysis to a one-time source-dependent computation. The key idea is: $ \mathrm{SSIM} \approx 1 - k \times \mathrm{MSE} $ where **k** captures the source-dependent characteristics of the image. Once computed, this single value can be reused across multiple encodes. ## The motivation Modern encoding workflows often evaluate many candidates: - encoder parameter searches - bitrate ladder optimization - per-title encoding - codec comparisons In these scenarios, repeatedly computing full SSIM can become a significant cost. The goal is not to replace SSIM, but to make SSIM-like quality estimation practical in workflows where many encodes need to be compared. ## The approach Previous work showed that local SSIM can be approximated from local MSE and source variance. This work extends that idea by estimating local distortion from the global frame MSE. The distortion is redistributed according to source activity, with higher-variance regions receiving a larger share of the error. Experiments show that a standard-deviation-based model provides a simple and effective approximation, allowing the local analysis to collapse into a single source-dependent factor. # Related Material - [QoMEX 2026 Poster](https://files.trud.ca/qomex2026-poster.pdf) - <img src="https://cdn.simpleicons.org/github" width="16" height="16"> [ApproxSSIMate on Github](https://github.com/luctrudeau/approxSSIMate) ## Acknowledgments Thanks to Maria G. Martini for the collaboration and for the previous work on the relationship between SSIM and PSNR/MSE that provided the foundation for this research.