Calibration of Structural Similarity Index Metric to Detect Artefacts in Game Engines

Abstract

Previous studies reveal that Image Quality Metics (IQMs) can be efficiently used to automatically detect perceptual visibility of artefacts in the game engines. Very good matching was achieved for shadow acne, peter panning, and Z-fighting deteriorations, while IQM with the best detection rate proved to be the Structural Similarity Index Metric (SSIM). However, this metric generates noticeably worse results for the aliasing. Using SSIM, the artefacts are identified as differences in intensity, contrast, and structure between an image with deterioration and the corresponding reference. In this work we calibrate SSIM to improve matching for aliasing artefacts. We compare results generated by SSIM with the reference data created during subjective experiments in which people manually mark the visible local artefacts in the screenshots from game engines. In other words, we maximise convergence in the detection between the maps created by humans and computed by SSIM. The results of the cross-validation performed on a large collection of examples revealed that AUC (area under curve) in the receiver-operator analysis can be improved from 0.92 for default SSIM parameters to 0.97 for optimised parameters.