New Mine-JEPA self-supervised learning approach outperforms billion-parameter foundation models at detecting underwater mines in sonar imagery using just 1,170 unlabeled images.

Mine-JEPA uses SIGReg regularization-based SSL loss to pretrain on only 1,170 unlabeled side-scan sonar images, addressing extreme data scarcity in maritime vision

Achieves F1 score of 0.935 for binary mine vs. non-mine classification, beating fine-tuned DINOv3 (0.922) which was pretrained on 1.7 billion images

Reaches 0.820 F1 score for 3-class mine-like object classification with synthetic data augmentation, still outperforming fine-tuned DINOv3 at 0.810

First in-domain SSL pipeline specifically designed for side-scan sonar, addressing the large domain gap between sonar data and natural images that foundation models struggle with