Introduction
Deep learning has revolutionized computer vision, yet its application in the underwater domain is severely constrained by a "data hunger" problem. Collecting and labeling sonar data is exceptionally challenging due to high operational costs, military confidentiality, and the inherent difficulty for humans to interpret low signal-to-noise ratio acoustic images. Unlike standard RGB photography, sonar images are characterized by acoustic shadows, multipath interference, and significant noise. This paper investigates the efficacy of Self-Supervised Learning (SSL) to mitigate the reliance on massive hand-labeled datasets for sonar image classification.
Methodology
The core approach involves using "pretext tasks" where the data itself provides the supervision signal. The researchers evaluated three primary SSL techniques to determine their effectiveness in the sonar domain:
- RotNet: The model is tasked with predicting the rotation (0°, 90°, 180°, or 270°) applied to an image. This forces the network to learn the natural orientation and structural features of underwater objects.
- Denoising Autoencoders (DAE): The network must reconstruct original sonar images from versions corrupted by Gaussian noise, learning to distinguish underlying object features from random acoustic interference.
- Jigsaw Puzzle: An image is divided into a 3x3 grid and shuffled; the model must predict the original permutation. This task emphasizes spatial relationships and local feature recognition across different parts of an object.
Following pre-training, the learned features are evaluated on downstream classification tasks using a Support Vector Machine (SVM) on separate datasets.
Experimental Setup
The study utilized three distinct datasets to test generalization and label efficiency:
- Marine Debris Watertank: 2,627 forward-looking sonar (FLS) images of 11 object classes (e.g., bottles, cans, tires) captured in a controlled environment.
- Marine Debris Turntable: 2,471 images of 12 classes captured at various angles on a rotating platform, serving as the primary test bed for transfer learning.
- "Wild" Sonar Dataset: A new, uncurated set of 63,000 patches extracted from sonar scans of a lake seabed. This dataset contains no class labels and represents realistic, messy underwater environments.
Results and Discussion
The experimental results highlight several key insights regarding the viability of SSL in underwater robotics:
- Label Efficiency: SSL pre-training achieved classification performance nearly identical to supervised methods. The Jigsaw model pre-trained on "Wild" data reached 97.02% accuracy on the Turntable dataset, nearly matching the supervised baseline of 97.09%.
- Robustness of Jigsaw: Among the three methods, the Jigsaw task proved most effective. Interestingly, the Jigsaw feature extractor is remarkably lightweight, utilizing only ~6,000 parameters compared to the ~1 million parameters in the RotNet and DAE architectures.
- Generalization from Uncurated Data: Remarkably, models pre-trained on "Wild" seabed patches—which contain no specific objects of interest—performed exceptionally well. This suggests that models can learn the fundamental "language" of sonar textures and shapes from raw environmental data.
- Few-Shot Performance: In scenarios with limited data (only 10 samples per class), SSL models provided a significant performance boost, outperforming the baseline by approximately 7%.
- RotNet's Limitations: RotNet failed significantly when trained on "wild" data (dropping to ~12% accuracy), likely because seabed patches lack a clear orientation, making rotation prediction impossible without distinct objects.
Conclusion
Self-supervised learning offers a viable path forward for underwater perception, significantly reducing the bottleneck of manual data labeling. While sonar features appear harder to generalize than RGB features—requiring more samples per class to reach peak performance—the success of Jigsaw-based pre-training on uncurated data demonstrates the potential for deploying robust perception systems on autonomous underwater vehicles (AUVs). Future research directions include exploring Transformer-based architectures and multi-modal SSL combining sonar with optical data.
Summary updated