English

中国农机化学报

中国农机化学报 ›› 2024, Vol. 45 ›› Issue (9): 258-264.DOI: 10.13733/j.jcam.issn.2095-5553.2024.09.039

• 农业智能化研究 • 上一篇    下一篇

基于 YOLOv4的水下海参检测与计数算法

宋小鹿 1,冯娟 1,2,梁翔宇 1,周玺兴 3   

  1. (1.河北农业大学信息科学与技术学院,河北保定,071001;2.河北省农业大数据重点实验室,河北保定,071001;3.河北农业大学机电工程学院,河北保定,071001)
  • 出版日期:2024-09-15 发布日期:2024-09-04
  • 基金资助:
    河北省重点研发计划项目(20327217D)

Recognition and counting algorithm of underwater sea cucumbers based on YOLOv4 network 

Song Xiaolu1,Feng Juan1,2,Liang Xiangyu1,Zhou Xixing3   

  1. (1. College of Information Science and Technology,Hebei Agricultural University,Baoding,071001,China; 2. Hebei Key Laboratory of Agricultural Big Data,Baoding,071001,China; 3. College of Mechanical and Electrical Engineering,Hebei Agricultural University,Baoding,071001,China) 
  • Online:2024-09-15 Published:2024-09-04

摘要:

针对智慧水产养殖中海参自动采捕和高效计量应用需求,提出一种基于 YOLOv4的水下海参检测与计数算法。该算法利用暗通道先验算法对数据集进行预处理,增强图像数据的可检测性;采用迁移学习方法训练 YOLOv4网络,并用 Swish函数替换骨干网络中的激活函数,提升自建数据集的海参检测性能;提出基于相近帧目标质心定位偏移的降重计数方法,优化目标计数结果。试验结果表明:该检测算法识别水下海参目标的平均检测精度的平均值 mAP达 91. 0%,分别比原始 YOLOv4、YOLOv3、Faster R-CNN和 SDD高 4. 5%、6. 9%、5. 0%、29. 9%;降重计数算法获得海参数量与人工计数结果间的均方根误差 RMSE为 29. 8、平均计数精度 ACP为 95. 8%、决定系数 R2为 0. 998。

关键词: 海参, 暗通道先验, YOLOv4, 迁移学习, 降重计数

Abstract:

In order to meet the requirements of automatic harvesting and high.efficiency measurement of sea cucumbers in intelligent aquaculture,a recognition and counting algorithm of underwater sea cucumbers based on YOLOv4 is proposed in this paper. The algorithm preprocesses the data set by using dark channel prior defogging algorithm to enhance the detectability of image data,YOLOv4 Network is trained with the transfer learning method,and Swish function is used to replace the activation function in the backbone network to improve the detection performance of the self-built data set,a method based on the target centroid positioning offsets of adjacent frames is proposed to optimize target counting result. The experimental results show that the mAP of sea cucumber targets recognized by the algorithm of this paper reaches 91. 0%, which is 4. 5%、6. 9%、5. 0% and 29. 9% higher than that recognized by original YOLOv4,YOLOv3,Faster R-CNN,and SDD,respectively. The RMSE between the number of sea cucumbers obtained by reducing repeated counts and the manual counting result is 29. 8. The average counting precision(ACP)is 95. 8% and the coefficient of determination(R2)is 0. 998.

Key words: sea cucumber, dark channel prior, YOLOv4, transfer learning, reducing repeated counts

中图分类号: