Object Detection in Small-Sample Underwater Sonar Images Using Improved YOLOv11

Underwater object detection in sonar images is challenging because of low image resolution, acoustic noise, seabed clutter, weak target boundaries, substantial target-scale variation, and limited labeled data. To address these challenges, we develop an improved You Only Look Once version 11 (YOLOv11)-based framework for small-sample sonar object detection. Rather than introducing new individual network components, the proposed framework coordinates hierarchical feature extraction, local contextual feature processing, and scale-adaptive detection according to the characteristics of sonar imagery. Specifically, a ResNet50 backbone is used to obtain deeper multilevel representations, a RepViT-enhanced C3k2 module is incorporated to strengthen the representation of small and low-contrast targets in cluttered backgrounds, a dual-pooling operation is employed to preserve prominent local echo responses and broader contextual information, and Dynamic Head is used to improve feature aggregation across different target scales and spatial locations. In addition, data augmentation and image synthesis techniques are applied to increase sample diversity and mitigate overfitting under limited-data conditions. Experiments on the Sonar Object Detection (SOD) and Underwater Acoustic Target Detection (UATD) datasets demonstrate the effectiveness of the proposed framework. On the SOD dataset, the model achieves an F1-score of 92.61%, an [email protected] of 77.4%, and an [email protected]:0.95 of 59.3%. On the UATD dataset, it achieves an F1-score of 84.0%, an [email protected] of 85.1%, and an [email protected]:0.95 of 38.1%. The experimental results confirm the model’s robustness in challenging underwater scenarios and its strong potential for high-precision applications in oceanic exploration, maritime rescue, and surveillance.