Individual Differences in Subjective Image Quality Assessment: A Bayesian Mixed-Effects Modeling Approach

Traditional approaches to subjective image quality assessment often rely on Mean Opinion Scores (MOS), which overlook individual differences among observers. This study addresses this limitation by introducing a novel experimental paradigm and Bayesian mixed-effects modeling framework, specifically Cumulative Link Mixed Models (CLMMs), to quantify these differences. A balanced dataset of 256 high-resolution images across four content categories, each with three color-based distortions, was rated by 16 observers. Our analysis demonstrates that models accounting for individual-level preferences outperform group-level models, offering more accurate predictions of subjective image quality ratings. The best-performing model accounted for both content and distortion preferences, highlighting the critical role of observer variability in image quality assessment. A follow-up experiment conducted one week later revealed trait-like stability in observers’ preferences for content and distortion types, with medium-to-large correlations in ratings (r = 0.43 to 0.97). These results indicate that individual differences are enduring traits rather than transient states. This study highlights the importance of considering human variability in subjective assessments, with significant implications for image quality methodologies and content delivery strategies. By embracing individual differences as valuable information, we can improve user experience, develop personalized content delivery systems, optimize compression techniques, and create more accurate predictive models for image quality. This work lays the foundation for a more individualized approach to image quality assessment, with applications in digital media, telecommunications, and human-computer interaction.