2025_programme: Procedural Seafloor Texture Generation for ATR training using synthetic data

Back


  • Day: June 17, Tuesday
      Location / Time: D. CHLOE at 12:40-13:00
  • Last minutes changes: -
  • Session: 18. Towards Automatic Target Recognition. Detection, Classification and Modelling
    Organiser(s): Johannes Groen, Yan Pailhas, Roy Edgar Hansen, Narada Warakagoda
    Chairperson(s): Johannes Groen, Yan Pailhas
  • Lecture: Procedural Seafloor Texture Generation for ATR training using synthetic data [Invited]
    Paper ID: 2312
    Author(s): Ciaran Sanford, Edward Clark, Alan Hunter
    Presenter: Ciaran Sanford
    Abstract: The development of automatic target recognition (ATR) in underwater environments has enabled remote detection and classification of objects, leveraging high resolution imaging techniques. ATR requires vast amounts of training data, but well-labelled real-world imagery is scarce. A recent development in synthetic aperture sonar (SAS) image simulation at the University of Bath has facilitated rapid generation of realistic synthetic echo data, enabling the use of synthetic data as training data for ATR. \n The environment models used for simulation are important, as their level of realism can drastically affect the performance of ATR. However, the seafloor models used in these simulations are limited. Existing seafloor models generally use either computational fluid dynamics (CFD) software or spectral methods, with the former being too slow for synthetic data generation, and the latter unable to model localised variations or object-floor interaction. As a result, existing methods are incapable of producing realistic textural diversity at the speed required to train ATR.\n We present a method of procedural texture generation that uses stochastic cellular automata to model the movement of sediment across the seafloor. The method is fast, using probabilistic state changes to reduce computational load. We present results showing localised texture variations with object-floor interaction, and demonstrate how it may be applied to ATR training for classification of objects.
  • Corresponding author: Dr Ciaran Sanford
    Affiliation: University of Bath
    Country: United Kingdom

Back