MF-CLF: Multi-Feature Chrominance–Luminance Fusion for Blind Underwater Image Quality Assessment

Underwater images commonly exhibit blurring, color casts, and low contrast due to light attenuation and scattering in water. Although numerous underwater image enhancement (UIE) algorithms have been developed to improve the usability of underwater imaging systems, evaluating the performance of these algorithms remains challenging due to the lack of reference images. Thus, blind/no-reference (NR) underwater image quality assessment (UIQA) has emerged as a key research focus. While existing NR-UIQA methods based on luminance and chrominance cues have shown effectiveness, modeling these attributes separately ignores valuable information arising from their joint behavior, since underwater degradations often induce simultaneous changes in luminance and chrominance that cannot be reliably characterized by either attribute alone. In this paper, we propose a lightweight and explainable NR-UIQA method, called multi-feature chrominance–luminance fusion (MF-CLF), based on jointly modeling the intra- and cross-attribute dependencies among chrominance and luminance statistics. Specifically, our approach constructs chrominance-attribute features across multiple color spaces, extracts luminance-attribute features using multi-kernel perceptual descriptors, and models the chrominance–luminance characteristics by explicitly capturing the interactions between the luminance and chrominance attributes. The extracted features are then mapped into a quality score using a support vector machine (SVM), enabling objective and reliable underwater image quality prediction. Experimental results tested on four public benchmark datasets demonstrate that MF-CLF significantly outperforms among lightweight, statistical-learning-based methods. Specifically, our approach achieves an SROCC value of 0.864 on the SAUD2.0 dataset, outperforming existing methods by 20.3%, and demonstrates strong robustness in cross-dataset evaluations with an SROCC value of 0.737, which is more than twice that of the traditional methods.