Deep Learning in Shallow Water: CNN-based 3D-FLS Target Recognition

Automatic target detection is important for real time navigation applications of 3D forward looking sonar (3D-FLS) systems. A 3D-FLS sensor generates a large volumetric point cloud of data that updates on the order of seconds, so that manually interpreting the data is not feasible for a vessel operator. FarSounder has developed an algorithm for detecting two classes of targets (seafloor and in-water-targets) based on traditional image and signal processing techniques. In this work, two modified versions of CNN architectures previously developed for volumetric data (3D U-Net and 3D V-Net) are evaluated for their ability to replace the current detection algorithm. The current detection algorithm was used to generate the set of training data and validation data for training and evaluation. Additional models were developed based on the 3D U-Net and V-Net models to operate on 2D cross sections of the data volumetric data as input instead of the volumetric input. All of the volumetric models achieved higher validation and training accuracy than the 2D versions, and the 3D U-Net replicated the traditional algorithm most closely. Finally, an automated procedure for improving training data using NOAA bathymetry data is described.