Underwater target detection based on target feature enhancement and semantic-position path aggregation

Abstract

Underwater object detection remains a challenging task due to severe image degradation, scale variation, and frequent occlusions, which often result in high false and missed detection rates. To address these issues, this paper proposes a novel underwater target detection framework that integrates feature enhancement with semantic-spatial guided fusion. Built upon the RT-DETR architecture, the proposed model introduces three key innovations. First, a Multi-Scale Edge Feature Injection (MSI-Edge) module is designed to incorporate fine-grained edge information from shallow layers into deeper representations, significantly improving the detection of small and low-contrast targets. Second, a Global–Local Feature Enhancement (GLF-Enhance) module replaces conventional multi-head self-attention, enabling efficient and balanced learning of both global contextual and local structural information while reducing computational overhead. Third, a Semantic–Location Path Aggregation Network (SL-PAN) is proposed to enable bidirectional interaction between semantic and spatial features, effectively mitigating information degradation during multi-scale feature fusion.Extensive experiments on benchmark underwater datasets demonstrate that the proposed method consistently outperforms the baseline RT-DETR (ResNet50 backbone), achieving improvements of 3.2%, 3.0%, and 2.7% in AP, AP50, and AP75 on the URPC dataset, and 2.9%, 2.7%, and 3.0% on the DUO dataset, respectively. Moreover, the model significantly reduces false positives and missed detections while maintaining real-time performance. Comprehensive ablation studies further validate the effectiveness of each module. Overall, the proposed approach establishes a robust and efficient solution, advancing the state-of-the-art in underwater object detection.