RCDI-YOLO: a target-detection method for complex environment side-scan sonar images based on improved YOLOv8

Target detection in side-scan sonar images under complex environments is challenging due to noisy backgrounds, occlusion, and blurred target boundaries, which reduce the accuracy and robustness of traditional methods. To address these issues, we propose RCDI-YOLO, an enhanced YOLOv8-based detection framework that integrates rotation-aware feature extraction, multi-scale feature integration, and implicit feature representations for noise suppression. In addition, a diversified complex environment side-scan sonar dataset (CESSSD) is constructed to mitigate data scarcity and imbalance. Experimental results demonstrate that RCDI-YOLO achieves a detection accuracy of 95.3% and a mean Average Precision of 95.7%, outperforming the original YOLOv8 by 2.5% and 2.0%, respectively. These findings confirm that RCDI-YOLO significantly improves detection performance in complex underwater environments, particularly in scenarios with occlusion, cluttered backgrounds, and noise interference, highlighting its potential for underwater detection and search-and-rescue applications.