Prior-Based Underwater Enhanced Image Quality Assessment Network

Abstract

Underwater images generally suffer from color cast and haze effects due to light attenuation and scattering, which leads to image quality degradation and poor recognition of image content by autonomous machines. Most of the existing enhancement algorithms try to remove these distortions of underwater images but do not perform perfectly. Moreover, there is no quality evaluation metric that can accurately measure the quality of these enhanced results. Thus, accurately evaluating the enhanced image quality is one of the urgent problems to be solved in underwater imaging research. In this article, a prior-based underwater enhanced image quality assessment network is proposed to measure the quality of those enhanced images objectively. First, underwater imaging priors, including object–camera distance map, ambient light, absorption and scattering parameters, surface–object distance, etc., directly affect the degree of color cast and haze effect in underwater images. Since the underwater raw image is available in the image enhancement task, a novel prior estimation network is designed to estimate these prior parameters from underwater raw images and obtain reliable prior information. Second, a novel prior guidance module is designed to guide these prior features to the enhanced image quality assessment network by simulating the underwater physical model. Ultimately, the quality of the enhanced image can be accurately evaluated through the end-to-end network. Furthermore, experiments show that the prior information can make the quality assessment network pay more attention to the content and distortion of the image, so as to evaluate the quality of the enhanced image more accurately. Extensive experiments on authentic data sets demonstrate the superiority of our model against other representative state-of-the-art models in both quantitative and qualitative results.