Evaluation of Kolmogorov-Arnold-inspired networks for blind sharpness assessment of authentic images

Without prior information or knowledge of the pristine images, blind image sharpness assessment (BISA) aims to quantify blurriness distortions that may result from the improper acquisition settings, imperfect imaging devices, or inappropriate data processing algorithms. Score regression-based BISA achieves remarkable progress, yet it remains challenging, particularly when authentic distortions exist in images. Support vector regression and other regression algorithms are commonly used to construct the quantitative relationship between handcrafted features and corresponding ratings. Alamogordo-Arnold-inspired networks (KANs) have demonstrated significant success in numerical fitting. In this study, a Taylor expansion-based model (TaylorKAN) is described, and systematic evaluation of both traditional algorithms and KAN models is performed by using 15 mid-level features as image sharpness representations. The evaluation of 12 traditional algorithms and 6 KAN models conducted on four public authentic image databases suggests that the TaylorKAN outperforms the other KAN models and the traditional regression algorithms with closer performance to human subjective ratings. Review of state-of-the-art works highlights that KAN models have shown promise, while there is still room for further improvement in score regression and no-reference image quality assessment.