Enhanced underwater object detection with YOLO-LDFE: a model for improved accuracy with balanced efficiency

In underwater image analysis, challenges such as complex environments, low model performance, and slow processing efficiency hinder effective object detection, which is crucial for real-time monitoring. To address these issues, we propose a high-precision underwater object detection model named YOLO-LDFE. To overcome the limitations of existing datasets, we have developed a comprehensive multi-species underwater biological dataset, MCUA, containing 9327 labeled images across 14 categories. The primary contributions of this work are as follows: (1) SPPF_SCSA, which integrates multi-semantic information with channel-space attention mechanisms to enhance performance; (2) the substitution of traditional convolutions with GSConv in C2f, reducing model size while maintaining feature extraction; (3) LEDF, which improves performance through multi-level, dense connections. YOLO-LDFE achieves exceptional results, with an average precision of 93% on the URPC2021 dataset, outperforming existing algorithms while maintaining high detection speed, demonstrating its potential for real-time underwater monitoring.