Underwater image enhancement and marine snow removal for fishery based on integrated dual-channel neural network

Abstract

Computer vision technology can reduce the intensity and difficulty of marine aquaculture. Nonetheless, the degradation problems of marine fishery underwater images hamper further interpretation and analysis of underwater information by computer vision technology. In order to comprehensively solve the problems of degradation in marine fishery underwater images, this paper proposes an end-to-end integrated dual-channel network model, which uses the underwater image enhancement module based on the residual dense network to perform color correction and dehazing for the low-frequency layer of images, and the marine snow removal module based on local residual learning strategy to remove the white marine snow in the high-frequency layer of images. Moreover, the addition of refinement module further improves the textual details and colors of images. Experimental results on marine fishery dataset indicate that the approach proposed in this paper performs better than several state-of-the-art methods in quantitative metrics, including underwater image quality measure (UIQM), blur index and smooth index, effective in improving the contrast of underwater fishery images and reducing marine snow noise, while achieving better visual quality in qualitative evaluation such as color correction, dehazing, detail and feature restoration. Furthermore, the generalization tests and application tests have proved its effectiveness in underwater scenarios, which means it can meet the practical needs in the field of marine aquaculture.