EFE_UODNet: Enhanced Underwater Organism Detection in Complex Environments

The detection of underwater organisms represents a particularly challenging frontier within the field of computer vision. This paper proposes an underwater organism detection algorithm named the Enhanced Feature Extraction-based Underwater Organism Detection Network (EFE_UODNet). Firstly, this paper designs an Enhanced Global Context Guided Feature (EGCGF) module to extract organism feature information in a serial interaction manner, thereby enhancing the feature extraction capability for blurred organisms in low-quality underwater images. Secondly, this paper proposes an Advanced Multi-Scale Fusion Pyramid Network (AM-FPN) to achieve multi-level feature fusion. EFE_UODNet uses high-level features as weights and combines channel attention with spatial attention to fuse low-level information with high-level features. Finally, the Dynamic Head (DyHead) is introduced to better utilize the features output by AM-FPN, thereby improving detection performance in underwater scenes involving dense and highly similar targets. Experimental results on the DUO dataset demonstrate that the proposed EFE_UODNet model, using YOLOv8m as the baseline, achieves significant improvements across multiple metrics. These improvements include a 2.3% increase in Precision, 3.5% in Recall, 2.1% in F1-score, 3.5% in mAP[50–95], and 1.3% in mAP50. These accuracy gains are achieved while ensuring that inference time and parameter count still meet the real-time and resource requirements of this work. Additionally, its generalization ability is validated on the URPC2019 dataset. Compared to other underwater organism detection algorithms, the proposed model delivers outstanding overall performance on this task.