No-Reference Generic Image Quality Assessment via Adaptive Fusion of Naturalness-Enhanced Colour and Appearance-Driven Spatial Metrics
IEEE Access · 2026 · 14 · 79811-79832
DOI: 10.1109/access.2026.3694284 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Muhammad Usman Khan und 2 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
No-reference image quality assessment (NR-IQA) remains a challenging problem due to the diverse nature of image distortions and the absence of a reference image. While most existing no-reference image quality metrics (NR-IQMs) mainly target spatial distortions, comparatively few methods address appearance-driven colour renderings, and generalization across heterogeneous datasets remains limited. To address this gap, this study proposes a unified NR-IQA framework that adaptively integrates colour and spatial quality predictions through a distortion-aware fusion strategy. The framework comprises three metrics: 1) CIQMZN, an enhanced colour-domain image quality metric (IQM) that extends our earlier CIQMZ metric by incorporating five naturalness-related statistical and texture descriptors; 2) SIQMZ, a spatial-domain IQM that derives natural scene statistics (NSS) features from the chroma and depth channels of the CIELAB colour space; and 3) GIQMZ, a classifier-guided generic metric that adaptively fuses CIQMZN and SIQMZ predictions based on rendering-type probabilities. Extensive evaluations were conducted on 17,387 images from seven benchmark datasets using both same-database and cross-database protocols. The results demonstrate that CIQMZN significantly improves predictive capability over CIQMZ for both colour renderings and spatial distortions, SIQMZ achieves competitive performance on spatially distorted datasets with a compact feature set, and GIQMZ provides robust and generalizable performance across mixed-distortion datasets through distortion-aware probabilistic fusion. Although deep learning-based models achieve higher prediction accuracy on large-scale datasets, the proposed framework offers a favourable trade-off between performance, interpretability, and computational efficiency. The source code for this work is available at https://github.com/mani0888/GIQMZ
Zitieren
APA
Khan, M. U., Mehmood, I., & Luo, M. R. (2026). No-Reference Generic Image Quality Assessment via Adaptive Fusion of Naturalness-Enhanced Colour and Appearance-Driven Spatial Metrics. IEEE Access, 14, 79811-79832. https://doi.org/10.1109/access.2026.3694284
Vancouver
Khan MU, Mehmood I, Luo MR. No-Reference Generic Image Quality Assessment via Adaptive Fusion of Naturalness-Enhanced Colour and Appearance-Driven Spatial Metrics. IEEE Access. 2026;14:79811-79832. doi:10.1109/access.2026.3694284
BibTeX
@article{Khan2026,
author = {Khan, Muhammad Usman and Mehmood, Imran and Luo, Ming Ronnier},
title = {No-Reference Generic Image Quality Assessment via Adaptive Fusion of Naturalness-Enhanced Colour and Appearance-Driven Spatial Metrics},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {79811--79832},
doi = {10.1109/access.2026.3694284}
}
EUVIMED