基于改进YOLOv8的南极磷虾群智能识别技术研究

Research on Intelligent Identification Technology of Antarctic Krill Swarms Based on Improved YOLOv88

摘要: 南极磷虾作为全球渔业中资源量最大的生物，关系着极地渔业的捕捞效率和经济效益。南极磷虾的捕捞正逐步成为极地渔业发展的热点问题。随着水下声学探测技术与机器视觉的快速发展，精准且富有效率的声呐图像识别成为南极磷虾智能化捕捞的前沿技术。鉴于目前人工声呐图像识别工作时长较长、占用捕捞船船员资源和识别准确率及效率较低的缺点，本文提出YOLOv8s-CBAM模型架构。通过在YOLOv8s的主干网络和颈部嵌入CBAM (卷积块注意力模块)，实现了对复杂声呐背景的鲁棒性提升。利用平均/最大池化和通道–空间双维度动态加权，并结合C2f结构优化特征聚合效率。强化了模型对南极磷虾的特征响应能力。实验结果表明，YOLOv8s-CBAM在保持轻量级架构的同时，实现了mAP50达到0.92，识别速度超过30 FPS的性能平衡，在定量分析与定性分析优于YOLOv8及YOLOv8-BiFPN等常用模型。

Abstract: Antarctic krill, as the organism with the largest biomass in global fisheries, is intrinsically linked to the fishing efficiency and economic viability of polar fisheries. The harvesting of Antarctic krill is increasingly emerging as a focal issue in the development of polar fisheries. With the rapid advancements in underwater acoustic detection technology and machine vision, precise and efficient sonar image recognition has become a cutting-edge technology for intelligent krill harvesting. In light of the limitations of current manual sonar image recognition—such as prolonged processing times, substantial consumption of fishing vessel crew resources, and sub optimal accuracy and efficiency—this paper proposes the YOLOv8-CBAM model architecture. By integrating the CBAM into the backbone and neck of the YOLOv8s architecture, achieve enhanced robustness against complex sonar backgrounds. This integration utilizes average/max pooling along with channel-spatial dynamic weighting to reinforce the model’s feature response capabilities, specifically for Antarctic krill. The architecture maintains optimized feature aggregation efficiency by incorporating the C2f structure. Experimental results demonstrate that YOLOv8s-CBAM strikes an excellent balance between performance and lightweight architecture, achieving an mAP50 of 0.92 and a detection speed exceeding 30 FPS. Furthermore, both quantitative and qualitative analyses confirm its superior performance compared to popular baseline models such as standard YOLOv8 and YOLOv8-BiFPN.