Human-Centric Evaluation Framework for No-Reference Image Quality Assessment

Abstract

The rapid proliferation of digital imaging technologies has necessitated robust, objective methods for assessing image quality without reference standards. No-reference image quality assessment (NR-IQA) has emerged as a critical field, aiming to predict perceptual image quality based solely on the distorted content itself. This study proposes a human-centric evaluation framework for NR-IQA, integrating statistical modeling, perceptual feature extraction, and latent quality factor analysis to align algorithmic assessment with human visual perception. Leveraging theoretical constructs from natural scene statistics (Sheikh et al., 2005) and regression-based neural network approaches (Li et al., 2011), the framework synthesizes both spatial and frequency domain features to predict subjective quality scores. The methodology incorporates comparative evaluations against existing blind IQA techniques (Moorthy & Bovik, 2011; Mittal et al., 2013) and demonstrates improved correlation with human judgment across diverse distortion types. Key findings reveal that integrating human perceptual modeling significantly enhances NR-IQA reliability, offering a scalable and interpretable tool for automated quality assessment in real-world imaging systems. Limitations of current no-reference models, including sensitivity to high-level semantic distortions, are critically discussed, providing a foundation for future improvements.