FuIQA: Content-adaptive blind JPEG image quality assessment via multi-domain feature fusion

With the rapid proliferation of JPEG compression in digital communications, the demand for reliable No-Reference (NR) quality assessment has intensified. This paper proposes a content-adaptive blind JPEG image quality assessment technique via multi-domain feature fusion. The proposed framework addresses the limitation of existing metrics, which frequently misidentify complex scene textures as distortion. A comprehensive set of 20 candidate features is extracted, spanning frequency-spectral sparsity, spatial block artifacts, and natural scene statistics. To eliminate variables that bias the model toward image content, a Content-Adaptive Feature Selection (CAFS) mechanism utilizing Recursive Feature Elimination with Cross-Validation (RFE-CV) is deployed. The optimized feature subset is utilized to construct two distinct quality assessment models. The primary model, FuIQA, employs a feed-forward neural network tuned via Bayesian Optimization to capture complex non-linear mappings. In addition, a computationally efficient variant, FuIQA-Lite, is proposed using a linear regression framework to offer a transparent, closed-form mathematical solution ideal for resource-constrained edge devices. Experimental validation across four diverse datasets (Kodak, UCID, USC-SIPI, and Waterloo) confirms that both variants significantly outperform traditional evaluators and modern deep learning baselines such as NIMA. The models demonstrate robust cross-dataset generalizability; when trained on the Waterloo dataset, the non-linear FuIQA achieved an average blind cross-dataset R-squared of 0.9925 and a PLCC of 0.9962, while the linear FuIQA-Lite achieved an SROCC of 0.9808.