Underwater stereo imaging systems with single enhancement techniques

Abstract

Recently, underwater works such as underwater exploration and construction are actively adopting the unmanned automation technologies for work safety and efficiency. Visualization sensors are essential for underwater unmanned automation technologies to understand what happed under water. In this study, we investigated two kinds of underwater stereo imaging systems using the commercial underwater cameras and the watertight stereo camera. Since underwater optical images are suffered from low visibility and color casting, the stereo image pairs are enhanced by image processing-based image fusion and deep learning-based generative adversarial network (GAN) methods, cycle-consistent GAN and underwater GAN. And then stereo vision techniques are applied to reconstruct three-dimensional point clouds from the enhanced image pairs. We evaluated the enhancement performance in terms of underwater image quality measure and underwater color image quality evaluation. Also, the accuracies on color compensation and depth estimation were evaluated with the reconstructed point clouds. All single enhancement methods enhanced quantitatively and qualitatively the deteriorated color and details of underwater stereo data, however, they did not help to improve the accuracies of depth estimation of the stereo data. Single enhancement is required to compensate color accuracy of underwater stereo imaging.