2023_programme: Applying Mask R-CNN to detect and classify underwater targets from synthetic aperture sonar data

Back


  • Session: 15. Towards Automatic Target Recognition. Detection, Classification and Modelling
    Organiser(s): Johannes Groen, Yan Pailhas, Roy Edgar Hansen, Jessica Topple and Narada Warakagoda
  • Lecture: Applying Mask R-CNN to detect and classify underwater targets from synthetic aperture sonar data
    Paper ID: 2030
    Author(s): Lopera Tellez Olga, De Meyst Olivier
    Presenter: Lopera Tellez Olga
    Abstract: Although automatic target recognition (ATR) algorithms have been developed for and are applied in different remote-sensing imaging applications, object-recognition tasks in this maritime application are still performed almost exclusively by human operators. ATR in remote sensing experiences a remarkable progress primarily due to the availability of large-scale datasets and deep convolutional neural networks (CNN).Despite the fact that obtaining large-scale datasets in the underwater threat recognition domain remains a challenge, very promising results have been observed with deep CNN architectures.In this paper, we propose to use transfer learning to apply Mask R-CNN for automatic classification of sea-mines, with the aim of evaluating a small architecture to be embedded into an AUV. Data to train and validate this architecture are collected using an AUV-mounted SAS with a theoretical 3.0 cm x 3.5 cm resolution. The centre frequency of the SAS is 300 kHz, and the bandwidth is approximately 60 kHz.Testing data were collected during mine countermeasures trials conducted in 2021 and 2022 close to the coast of Portugal. The target class comprises exercise mines that were purposely deployed prior to the surveys; the clutter class comprises other man-made objects (deliberately deployed), rocks, seafloor anomalies and all other alarms.\nWe evaluate when and why transfer learning from pre-trained deep CNN can be useful. As CNN architectures are considered by end-users as black-box models, this study also proposes some interpretation of the learned filters and the intermediate responses, and how they could be liked to features related to the target (such as highlight and shadow).
  • Corresponding author: Dr Olga Lopera Tellez
    Affiliation: Signal and Image Centre, Royal Military Academy
    Country: Belgium
    e-mail:

Back