基于改进YOLOv5的海洋生物检测算法

Marine Organism Detection Algorithm Based on Improved YOLOv5

对于水下机器人捕捞行业,提高对海洋生物的识别准确率可以有效减少对海洋生态环境的破坏。针对现有海洋生物检测模型在复杂环境下对小目标存在特征提取能力不足、检测精度低的问题,提出了一种基于改进YOLOv5的海洋生物检测算法。该算法通过改进马赛克数据增强,生成更多的小目标数据样本;在YOLOv5的主干网络引入SimAM(Simple,Parameter-Free Attention Module)无参注意力机制,该注意力机制对海洋生物特征图分配3D注意力权值,从而增强模型提取特征的能力。实验结果表明,对比原始YOLOv5算法,在没有引入额外参数的情况下,查准率、查全率、平均检测精度分别提高了2.8%、1.7%、1.6%,为水下机器人进行精准识别和捕捞提供了技术支持。

For the underwater robot fishing industry,improving the recognition accuracy of marine organism can effectively reduce the damage to the marine ecological environment.For the existing marine organism detection model in a complex environment for small targets,there are problems such as insufficient feature extraction ability and low detection accuracy,an improved YOLOv5 target detection algorithm is proposed to detect the marine organism.The algorithm generates more small target data samples by improving the Mosaic data augmentation.A Parameter-Free Attention Module is added to the backbone network of YOLOv5,which assigns 3D attention weights to the marine organism feature map,so that the model’s ability to extract features is improved.The experimental results show that the precision,recall and average detection accuracy are improved by 2.8%,1.7%and 1.6%in comparison with original YOLOv5 algorithm,respectively,without additional parameters,which provides a technical support for accurate identification and fishing of underwater robots.