New math framework optimizes image denoising for human perception

A team led by Saara Isoranta and Emilia L.K. Blåsten from the University of Helsinki has published a paper (arXiv:2606.00122) that rethinks image denoising through the lens of human perception. Instead of relying solely on mathematical error metrics like PSNR or SSIM, they propose a framework that asks: which denoising parameters produce images that humans find most visually similar to the original? Their method uses a total variation denoising algorithm with a tunable parameter, generating a set of differently denoised versions of the same base image. They then conduct controlled human comparison tests to measure perceived dissimilarity. Using psychometric scaling, they convert subjective human judgments into a quantifiable HaarPSI (Haar Wavelet-Based Perceptual Similarity Index) value, which serves as a threshold for discretizing the parameter grid.

This work produces two key outputs: first, openly available, psychometrically scaled image sets that researchers can reuse; second, a generalizable framework for any comparison-based imaging experiment. The practical implication for computer vision and photography is significant: denoising algorithms can now be tuned not just to minimize noise statistically, but to align with what the human visual system actually prefers. This could improve everything from smartphone photography to medical imaging, where retaining diagnostic detail while making images look natural is critical. The team has released the resulting datasets and methodology for open use, inviting further experiments in perception-driven imaging.