Towards Robust Object Detection in Underwater Sonar Imagery: A Cross-Modality Transfer Learning Study Leveraging RGB, FLS, and SAS Data

Franziska Auer, Tobias Meisen, Heeraj Ayyappan

Keywords: synthetic aperture sonar, sonar target detection, yolo

TL;DR: This paper investigates cross‑modality transfer learning, targeted augmentations, and sensor‑aware dataset mixing to improve deep‑learning‑based object detection in scarce sonar imagery.

Abstract: Robust underwater object detection is challenged by the complex acoustic nature of sonar imagery and the scarcity of labeled data. This research investigates cross-modality transfer learning to address these issues. Given the scarcity of labeled synthetic aperture sonar (SAS) images, we explore the potential of leveraging the larger and more readily available labeled datasets from RGB and forward-looking sonar (FLS) to improve object detection performance in SAS images. We systematically analyze the impact of sensor differences, data augmentation techniques, and dataset mixing strategies. Our results demonstrate that pre-training on RGB datasets can significantly enhance SAS detection performance, particularly when utilizing convolutional neural networks. However, direct transfer learning is ineffective without sonar-specific adaptation, and combining datasets from disparate sonar sensors can be detrimental due to feature inconsistencies. These findings highlight the need for tailored approaches to data processing and model selection in sonar image analysis, paving the way for more robust and efficient underwater object detection systems.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 5