Deep No-Reference Quality Assessment for Underwater Enhanced Images

The goal of underwater image enhancement (UIE) is to increase the quality of acquired underwater images, which significantly increases the value of these images. However, without effective underwater enhanced image quality assessment (UEIQA) measures that benchmark UIE, the UIE process becomes driftless, and the enhanced results produced by different UIE algorithms cannot be fairly compared. To this end, in this work, we construct a dedicated UEIQA scheme on the basis of a deep investigation of the characteristics of enhanced underwater images. Specifically, in our proposed method, we design deep neural networks to represent the unique attributes of enhanced underwater images, such as color casts, local distortions, degrees of naturalness, sharpness levels, contrast levels, and fog densities, which are highly correlated with image quality. Then, we introduce a vision transformer (ViT) to capture the dependencies among different image attributes and infer the quality level of the examined images. Extensive experiments conducted on three typical UEIQA databases, i.e., SOTA, UID2021 and SAUD, show that the proposed UEIQA model yields notably higher prediction accuracy than do the representative IQA and UEIQA metrics, e.g., achieving SRCC values of 0.891 (vs. 0.749) on SAUD and 0.933 (vs. 0.798) on UID2021.