Decomposition of Perceptual Scores in Multi-Component IQA: Pathways to Enhanced Model Interpretability and Efficiency
Type: article
Abstract: Traditional Image Quality Assessment (IQA) often aggregates human perception into a single score, overlooking its multidimensional nature, which is crucial for interpretable IQA modeling.This study presents a systematic cross-dataset analysis of how technical distortions or aesthetic components jointly shape overall quality perception.By evaluating seven large-scale datasets (SPAQ, KonIQ-10k, GISET, ICAA20K, AADB, EVA, and PARA), we investigate the relationships between human-annotated attribute-level features and Mean Opinion Scores (MOS) using Pairwise Correlation (PC), Principal Component Analysis (PCA), and cross-validated regression models.Our analysis uncovers four distinct empirical patterns in perceptual score modeling: (1) high attribute redundancy (SPAQ, GISET, EVA, PARA), where inter-attribute correlations exceed 0.7-0.9,achieving near-optimal predictive power (PLCC/SRCC > 0.96) and indicating that existing multiattribute annotations often collapse into a single dimension, rendering complex architectural partitioning redundant for these specific tasks; (2) effective orthogonality in AADB, where aesthetic attributes show weak inter-correlations (𝑟 < 0.3) yet enable accurate regression (PLCC/SRCC ≈ 0.90), demonstrating that orthogonal features contribute unique variance; (3) an extraction and complexity gap in KonIQ-10k, where structurally simple, automated attributes fail to align with actual subjective perception (PLCC/SRCC ≈ 0.74); and (4) insufficient specificity in ICAA20K, where isolated perceptual color attributes (temperature, harmony) show weak correlation with MOS (PLCC ≈ 0.34), proving inadequate without semantic context.Based on these findings, we propose actionable design principles for future IQA benchmarks, prioritizing semantically distinct attributes alongside annotation protocols designed to mitigate cognitive biases (e.g., the halo effect), thereby ensuring interpretable and robust IQA models.
Source: CyberLeninK (CyberLeninka)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225030228