LoRA Enhanced Contrastive Learning with SAS Vision Transformers

基于 LoRA 增强对比学习的 SAS 视觉 Transformer 研究

Abstract: Automatic target recognition (ATR) with synthetic aperture sonar (SAS) supports advanced naval capabilities, but deep learning is constrained by scarce target imagery, background clutter, and human-in-the-loop assessment. We adapt DINOv3 Vision Transformer (ViT) models to underwater SAS ATR using a three-stage parameter-efficient framework.

摘要： 合成孔径声纳（SAS）的自动目标识别（ATR）技术支持先进的海军作战能力，但深度学习的应用受到目标图像稀缺、背景杂波以及人工评估环节的限制。我们采用了一个三阶段的参数高效框架，将 DINOv3 视觉 Transformer（ViT）模型适配到水下 SAS ATR 任务中。

Stage 1 uses Low-Rank Adaptation (LoRA) while freezing the ViT backbone, bridging the gap between natural-image pretraining and underwater acoustic propagation. Stage 2 uses hard-negative mining to strengthen the decision boundary against acoustic mimics, including rocks and sediment formations resembling man-made targets. Stage 3 uses Supervised Contrastive Learning (SupCon) to separate target and clutter representations.

第一阶段使用低秩自适应（LoRA）技术，同时冻结 ViT 主干网络，从而弥合自然图像预训练与水下声学传播之间的差距。第二阶段使用难例挖掘（hard-negative mining）来加强决策边界，以应对声学模拟物，包括那些类似于人造目标的岩石和沉积物结构。第三阶段使用监督对比学习（SupCon）来分离目标与杂波的特征表示。

We evaluate at-sea SAS data using a mission-level geographic split, compare all arms at 85 percent test recall, and repeat each comparison over three random seeds. LoRA accounts for the primary effect, increasing area under the precision-recall curve (AUPRC) from 0.300 to 0.679 +/- 0.027 using the same frozen backbone. Rank 4 achieves this result while training only 0.26 percent of weights.

我们使用任务级的地理划分对海上 SAS 数据进行了评估，在 85% 的测试召回率下比较了所有实验组，并针对三个随机种子重复了每次比较。LoRA 产生了主要效果，在使用相同冻结主干的情况下，将精确率-召回率曲线下面积（AUPRC）从 0.300 提升至 0.679 +/- 0.027。秩（Rank）为 4 的设置在仅训练 0.26% 参数的情况下实现了这一结果。

Neither refinement stage exceeds its matched control: hard-negative mining changes AUPRC by -0.0045 +/- 0.0119 versus an equal-size random curriculum, and SupCon changes AUPRC by +0.0002 +/- 0.0096 versus the preceding stage. These null results indicate that mining occurred on data the encoder had already fit and that supervised stages had already imposed most target-clutter geometry. One efficient adaptation stage is sufficient; stacked refinement is not.

后续的两个优化阶段均未超过其对照组的表现：难例挖掘相比同等规模的随机课程学习，使 AUPRC 变化了 -0.0045 +/- 0.0119；而 SupCon 相比前一阶段使 AUPRC 变化了 +0.0002 +/- 0.0096。这些无效结果表明，挖掘过程发生在编码器已经拟合的数据上，且监督阶段已经确立了大部分目标与杂波的几何关系。结论是：一个高效的适配阶段足矣，堆叠式优化并无必要。