Prototype Relation Embedding and Contrastive Learning for Improved Few-Shot Object Detection in Sonar Images

While recent advancements have significantly elevated the performance of object detection in sonar images, the performance of object detectors experiences a steep decline when trained with a limited number of samples. Few-shot object detection (FSOD) in sonar images remains a challenge. In this paper, we propose a novel method based on prototype relation embedding and contrastive learning for few-shot object detection named SRE-Net for sonar images. In the SRE-Net framework, prototype relation embedding network is employed to enhance class prototype representations, facilitating improved region proposal generation by utilizing weighted class prototypes computed from query images. Furthermore, the introduction of a support-guided detection head enables the model to exploit the associative information between query and support images, thereby strengthening inter-object relationships. Meanwhile, a contrastive head is introduced to enhance feature representations, reducing intra-class variations while preserving inter-class differences. Finally, we compare SRE-Net with state-of-the-art few-shot object detection methods on optical and sonar image datasets. The results show that our model demonstrates outstanding detection capability across various evaluation scenarios.