Article
In a groundbreaking study published on arXiv, researchers have developed an innovative framework for automatic target recognition (ATR) in underwater environments, utilizing synthetic aperture sonar (SAS) and advanced deep learning techniques. The paper, titled "LoRA Enhanced Contrastive Learning with SAS Vision Transformers", outlines a three-stage approach that leverages Low-Rank Adaptation (LoRA) to enhance the performance of DINOv3 Vision Transformer (ViT) models.
The challenge of ATR in underwater settings is compounded by the scarcity of target imagery and the presence of background clutter, making accurate identification difficult. The researchers address these issues by adapting the ViT models specifically for SAS ATR, starting with a stage that employs LoRA to bridge the gap between pretraining on natural images and the unique characteristics of underwater acoustic propagation.
Subsequent stages of the framework incorporate hard-negative mining to improve the model's decision-making against misleading acoustic mimics, such as rocks and sediment that may resemble man-made targets. The final stage employs Supervised Contrastive Learning (SupCon) to further differentiate between actual targets and clutter. The results from at-sea SAS data demonstrate a significant increase in the area under the precision-recall curve (AUPRC), showcasing the effectiveness of the LoRA adaptation.
Notably, the study found that LoRA alone was responsible for a remarkable increase in AUPRC from 0.300 to 0.679 while only training a mere 0.26 percent of the model's weights. This efficiency suggests that a single adaptation stage can be sufficient for effective performance, challenging the necessity for multiple refinement stages. Such findings are crucial for enterprises involved in AI product development, particularly those focused on enhancing recognition capabilities in complex environments.