Machine Learning for Inferring Depth from Side-scan Sonar Images

Sammanfattning: Underwater navigation using Autonomous Underwater Vehicles (AUVs), which is significant for marine science research, highly depends on the acoustic method, sonar. Typically, AUVsare equipped with side-scan sonars and multibeam sonars at the same time since they both have their advantages and limitations. Side-scan sonars have a much wider range than multibeamsonars and at the same time are much cheaper, yet they could not provide accurate depth measurements. This thesis is aiming at investigating if a machine-interpreted method could beused to translate side-scan sonar data to multibeam data with high accuracy so that underwater navigation could be done by AUVs equipped only with side-scan sonars. The approaches considered in this thesis are based on Machine Learning methods, including generative models and discriminative models. The objective of this thesis is to investigate the feasibility of machine learning based models to infer the depth based on side-scan sonar images. Different models, including regression and Generative Adversarial Networks, are tested and compared. Different CNN based architectures such as U-Net and ResNet are tested andcompared as well. As an experiment trial, this project has already shown the ability and great potential of machine learning based methods extracting latent representations from side-scansonars and inferring the depth with reasonable accuracy. Further improvement could be madeto improve the performance and stability to be potentially verified on the AUV platforms inreal-time.