2023_programme: Automatic target recognition using deep learning on multiple SAS views

• Session: 08. Munition Detection- Classification and Localization in the Marine Environment
  Organiser(s): David Bradley
• Lecture: Automatic target recognition using deep learning on multiple SAS views
  Paper ID: 2025
  Author(s): Bryan Oscar, Fincham-Haines Tom, Hunter Alan, Warakagoda Narada, Hansen Roy
  Presenter: Bryan Oscar
  Abstract: Modern autonomous vehicles combined with synthetic aperture sonar are capable of imaging many kilometres of seafloor in cm resolution. Due to the substantial amount of data generated, automating the detection and classification of seafloor objects becomes essential. This work presents a technique for combining multiple views of an object to train a deep learning model without the need for manual labels. By synthesising unseen views the model is able to learn the relationships to geometry and imaging physics. This has two benefits: first, the model’s interpretation of an input can be better understood by generating and inspecting novel views. Second, this model can be fine-tuned on a small amount of labelled data to better perform classification and detection. We demonstrate improved classification accuracy following the “re-view” pretraining compared to no pretraining, and pretraining using an autoencoder. We also show qualitative re-viewing results, demonstrating the improved interpretability from this machine learning approach.
• Corresponding author: Mr Oscar Bryan
  Affiliation: University of Bath
  Country: United Kingdom
  e-mail: