LoRA Improves Underwater Sonar Target Recognition with DINOv3 ViTs

Summary

The paper studies automatic target recognition in synthetic aperture sonar (SAS), where scarce imagery, background clutter, and human assessment constrain deep-learning systems. It adapts pretrained DINOv3 Vision Transformer models through a three-stage parameter-efficient framework. The first stage applies Low-Rank Adaptation (LoRA) while freezing the backbone; the second mines hard negatives such as rocks and sediment that resemble man-made targets; and the third applies Supervised Contrastive Learning to separate target and clutter representations. Evaluation uses at-sea SAS data with a mission-level geographic split, compares methods at 85% test recall, and repeats each comparison across three random seeds. LoRA produces the main improvement, raising area under the precision-recall curve from 0.300 to 0.679 +/- 0.027 with the same frozen backbone. Rank 4 achieves this result while training only 0.26% of the model weights. Neither added refinement improves its matched control: hard-negative mining changes AUPRC by -0.0045 +/- 0.0119, while supervised contrastive learning changes it by +0.0002 +/- 0.0096. The authors interpret these null results as evidence that the encoder had already fitted the mined examples and that the supervised stages had already established most target-clutter geometry, making one efficient adaptation stage sufficient for this setting.