Semi-supervised advancement of underwater visual quality

Abstract

Abstract In the underwater environment, the backscattering and attenuation of wavelength-dependent light degrade the quality of underwater vision. Low-quality underwater vision will reduce the accuracy of underwater robot visual navigation and pattern recognition. A novel semi-supervised deep convolutional neural network composed of a supervised learning branch and an unsupervised learning branch is proposed herein to improve underwater visual quality with poor visibility in real time. The network is constrained by a supervised loss function consisting of mean square, underwater index, and adversarial loss. The supervised branch serves as the baseline of the image enhancement algorithm to learn the basic feature information of the images and restore the original colors. The unsupervised learning branch, which makes the generated images more realistic and reduces reliance on the quality of the simulation model of synthetic data, applies underwater dark channel prior loss and total variation loss to learn the feature domain information of real images. Experiments show that the results of the proposed method show less color shift, lower fogging and blurring, and more pleasing high-quality vision. The enhanced images can extract more useful feature information, which is promising in the online visual navigation of underwater robots.