English

中国农机化学报

中国农机化学报 ›› 2024, Vol. 45 ›› Issue (7): 228-234.DOI: 10.13733/j.jcam.issn.2095-5553.2024.07.034

• 农业智能化研究 • 上一篇    下一篇

基于多路神经网络的黄花菜关节点检测算法研究

张延军1, 2,赵建鑫1, 2   

  1. 1. 太原科技大学机械工程学院,太原市,030024; 
    2. 山西省重大装备液压基础元件及智能制造工程研究中心,太原市,030024
  • 出版日期:2024-07-15 发布日期:2024-06-24
  • 基金资助:
    国家自然科学基金(62001321);山西省重点研发计划项目(201903D121041)

Research on daylily joint detection algorithm based on multiple neural networks 

Zhang Yanjun1, 2, Zhao Jianxin1, 2   

  1. 1. School of Mechanical Engineering, Taiyuan University of Science and Technology, Taiyuan, 030024, China;
    2. Research Center Basic Hydraulic Components and Intelligent Manufacturing Engineering of 
    Major Equipment, Taiyuan University of Science and Technology, Taiyuan, 030024, China
  • Online:2024-07-15 Published:2024-06-24

摘要: 传统的目标检测算法仅能得到目标框,无法确定黄花菜的生长姿态。针对这一问题,在现有目标检测算法的基础上优化神经网络(Neural Network)结构,由检测框预测改为关节点的预测。首先,按照锚点匹配的方式确定黄花菜的生长方向及长度,统计黄花菜目标生长角度和长度,按照统计结果设置多个锚点,实际生长角度和长度与锚点进行比较,获得目标的相对长度和角度,并将其作为模型预测值进行训练;其次,在模型中加入热力图预测分支,对4个关节点进行预测;最后,利用目标长度和角度信息连接关节点得到黄花菜目标的生长姿态。设计针对线段拟合特点的评估模型方法,在计算精度的过程中引入部分亲和度字段。并据此改进非极大值抑制算法(NonMaximum Suppression)。试验结果表明:采用热力图校准后的模型对采摘目标识别精度达91.02%,定位精度达99.8%。

关键词: 黄花菜, 采摘机器人, 神经网络, 机器视觉, 图像处理, 目标检测

Abstract: The traditional target detection algorithm can only get the target frame, and cannot determine the growth direction of daylily. Aiming at this problem, the neural network structure is optimized on the basis of the existing target detection algorithm, and the prediction of the detection box is changed to the prediction of the key points. Firstly, the growth direction and length of daylily are determined according to the anchor point matching method, and the growth angle and length of the daylily target are counted. Multiple anchor points are set based on the statistical results. The actual growth angle and length are compared with the anchor points to obtain the relative length and angle of the target, which is used as the model prediction value for training. Secondly, the heat map prediction branch is added to the model to predict the four key points. Finally, the growth posture of daylily target is obtained by using the target length and angle information to connect the key points. An evaluation model method for line segment fitting characteristics is designed, Introduction of Partial Affinity Fields in Calculation Accuracy, and the NonMaximum Suppression algorithm is improved accordingly. Through experimental verification, the accuracy of picking target recognition is 91.02%, the positioning accuracy is 99.8%.

Key words: daylily, picking robot, neural network, machine vision, image processing, object detection

中图分类号: