Enhanced YOLOv8 Architecture for Object Detection in Sonar Imagery With Improved Localization Accuracy

This paper introduces two enhanced YOLOv8-based models, SPD-LKA-YOLO4H and SPD-LKA-YOLO4H-CBAM, to address key challenges in sonar imagery object detection, such as low visibility and small object detection. SPD-LKA-YOLO4H integrates Space-to-Depth convolution (SPD-Conv) to preserve fine-grained details and Large Kernel Attention (LKA) to enhance global and local feature representations. A further refined variant, SPD-LKA-YOLO4H-CBAM, incorporates a Convolutional Block Attention Module (CBAM) to refine the feature selection and suppress noise. A high-resolution detection head is also included to improve small object detection. The extensive evaluation on a combination of controlled and challenging real-world sonar datasets demonstrated consistent performance improvements. Specifically, the SPD-LKA-YOLO4H model achieved superior performance on a real-world dataset, demonstrating its generalization ability with a strong balance of accuracy and speed. The SPD-LKA-YOLO4H-CBAM model, although slightly slower, achieved a 13.21% gain in mAP@50–95 on the complex UATD dataset, making it ideal for tasks requiring high precision. The study also highlighted that the efficacy of these enhancements is highly sensitive to module placement, underscoring the critical role of architectural optimization. Overall, the proposed models strike an excellent balance between detection precision and computational efficiency, positioning them as promising solutions for real-time autonomous underwater applications.