Underwater Object Classification with Deep Learning

Measurement technologies for acquiring 3D shape data have advanced remarkably in recent years: point cloud surveying using laser pulse reflection on land, and acoustic beam technologies underwater, are both developing rapidly. Meanwhile, inspection of port facilities has mainly relied on visual checks by divers, but the number of workers has dropped sharply due to aging and safety concerns. Acoustic surveying using autonomous underwater robots is attracting attention as a solution. However, the resulting data volume is enormous, and because it contains no color information, even experts have struggled to interpret it reliably. We therefore developed a method that uses deep learning to classify and colorize point clouds by structure, making underwater objects much easier to identify.