Benthos-DETR：A High-precision Efficient Network for Benthic organisms Detection

Weibo Rao¹Gang Chen^1*Yifei Zhang²Jue Cang³Shusen Chen¹Chenyang Wang¹

• ¹College of Marine Science and Technology, China University of Geosciences Wuhan, Wuhan, China
• ²Hubei Institute of Water Resources Survey and Design CO..LTD, Wuhan, China
• ³Bureau of Hydrology Tibet, Lhasa, China

The intelligent, automated, and high-precision detection of underwater targets represents a challenging yet pivotal issue in marine science. Enhancing the localization accuracy of marine organisms holds significant importance for marine scientific research fields such as ecological conservation and fisheries management, especially in complex seabed environments where accurately identifying benthic organisms characterized by small size, large quantities, and diverse species offers considerable economic benefits and practical value. This study proposes Benthos-DETR, a benthic organisms detection network based on the RT-DETR network. In the backbone of Benthos-DETR network, the Efficient Block with the C2f module reinforces the shallow feature extraction operation in Benthos-DETR, enhancing the algorithm's multi-scale perception. To reduce the computational load and make the algorithm lightweight, a cascaded group attention module has been added to the Benthos-DETR network, it enhances the feature interaction within the same scale. In the neck, the original concatenation module is replaced with the Fusion Focus Module, effectively aggregating feature layer information from different stages of the backbone to achieve cross-scale feature fusion. The proposed Benthos-DETR ensures high target detection accuracy while minimizing hardware requirements for network deployment. The outcomes of the ablation experiment revealed that the various modules introduced in this research optimize the baseline network, and their integration markedly elevates the performance of Benthos-DETR. In tests on an open-source dataset, Benthos-DETR achieved a detection accuracy of 92.1% and mAP50 of 91.8% for sea cucumbers, 91.6% accuracy and 92.2% mAP50 for sea urchins, and 92.4% accuracy and 93.7% mAP50 for scallops. Through a series of experimental analyses, it was evident that the performance of the Benthos-DETR network surpasses existing target detection algorithms, achieving an optimal equilibrium between high recognition precision and a trim network scale.