Robust Unsupervised Cleaning of Underwater Bathymetric Point Cloud Data

This paper presents a novel unified one-stage unsupervised learning framework for point cloud cleaning of noisy partial data from underwater side-scan sonars. By combining a swath-based point cloud tensor representation, an adaptive multi-scale feature encoder, and a generative Bayesian framework, the proposed method provides robust sonar point cloud denoising, completion, and outlier removal simultaneously. The condensed swath-based tensor representation preserves the point cloud associated with the underlying three-dimensional geometry by utilizing spatial and temporal correlation of sonar data. The adaptive multi-scale feature encoder identifies noisy partial tensor data without handcrafted feature labeling by utilizing CANDECOMP/PARAFAC tensor factorization. Each local embedded outlier feature under various scales is aggregated into a global context by a generative Bayesian framework. The model is automatically inferred by a variational Bayesian, without parameter tuning and model pre-training. Extensive experiments on large scale synthetic and real data demonstrate robustness against environmental perturbation. The proposed algorithm compares favourably with existing methods.