LoRA-Enhanced Contrastive Learning Boosts Underwater Sonar Target Recognition
Summary:A new study on arXiv introduces a framework that combines Low-Rank Adaptation (LoRA) with contrastive learning to enhance Automatic Target Recognition (ATR) in underwater Synthetic Aperture Sonar (SAS) imagery. By freezing the DINOv3 Vision Transformer (ViT) backbone and applying LoRA for efficient parameter adaptation, the method leverages hard-negative mining and Supervised Contrastive Learning (SupCon) to improve target-clutter separation. Experiments demonstrate a significant increase in AUP
Background and Challenges
Underwater Synthetic Aperture Sonar (SAS) target recognition is crucial for advanced naval capabilities, but its deep learning applications face challenges such as scarce target imagery, background clutter, and inefficient human-in-the-loop assessment.
Methodology and Innovation
This study proposes a three-stage parameter-efficient framework that adapts the DINOv3 Vision Transformer (ViT) model for underwater SAS ATR:
- Low-Rank Adaptation (LoRA) Stage: While freezing the ViT backbone, LoRA is applied for efficient parameter adaptation, bridging the gap between natural-image pretraining and underwater acoustic propagation.
- Hard-Negative Mining Stage: Hard-negative mining strengthens the decision boundary against acoustic mimics, such as rocks and sediment formations resembling man-made targets.
- Supervised Contrastive Learning (SupCon) Stage: SupCon is used to separate target and clutter representations.
Experiments and Results
The study evaluates at-sea SAS data using a mission-level geographic split and compares all methods at 85% test recall, repeating each comparison over three random seeds. The results show that LoRA is the primary effect, increasing the Area Under the Precision-Recall Curve (AUPRC) from 0.300 to 0.679 ± 0.027 using the same frozen backbone. Training only 0.26% of the weights achieves this result.
Conclusion and Implications
The refinement stages of hard-negative mining and SupCon did not exceed their matched controls, indicating that mining occurred on data the encoder had already fit and that supervised stages had already imposed most target-clutter geometry. Therefore, one efficient adaptation stage is sufficient; stacked refinement is not.
Industry Impact and Developer Recommendations
- Underwater Sonar Applications: This method provides a new technical path for underwater sonar target recognition, especially in scenarios with scarce data and complex backgrounds.
- Parameter-Efficient Adaptation: The application of LoRA demonstrates its advantage in maintaining high performance while significantly reducing computational costs, providing a reference for deep learning applications in resource-constrained environments.
- Contrastive Learning Applications: The combination of hard-negative mining and SupCon offers new ideas for complex target recognition tasks in other fields.
— END —Source: ArXiv AI (cs.AI) (2026-09-21)
Tags: #LoRA #Contrastive Learning #Underwater Sonar #SAS #ViT
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments