Self-supervised pretraining-based zero-shot target detection for synthetic aperture sonar

The severe scarcity of annotated data in synthetic aperture sonar imagery significantly constrains the performance of automatic detection for underwater small targets. To address this issue, this paper proposes a zero-shot detection method based on deep learning. The core of this approach lies in leveraging large-scale self-supervised pre-training to overcome the dependency on annotated data: initially, a substantial volume of unlabeled sonar images is utilized to learn generic sonar feature representations through self-supervised pre-training; subsequently, the model is fine-tuned with a minimal set of annotated samples to equip it with zero-shot detection capability for small targets. Experiments conducted on a sonar dataset collected from the Jiamusi waters demonstrate that the proposed method achieves high-precision and highly robust autonomous identification of underwater small targets even under conditions of sample scarcity.