摘要
针对传统的目标检测方法依赖人工提取特征,存在检测效率低、鲁棒性差和实时性差等缺陷。本文根据刺梨果实在自然环境中的生长特点,采用带有残差模块的YOLO v3神经网络进行刺梨果实识别模型的训练,该网络通过提取不同卷积层的特征图,将深层特征图进行上采样后与浅层特征图进行多次融合,以提取图像的更深层次的特征信息。通过对该网络的相关参数进行优化和改进,并对未参与模型训练的70幅刺梨图像进行检测,实验表明,本文算法能够有效地对自然环境下的11类刺梨果实进行识别,各类识别平均准确率为88.5%,平均召回率为91.5%,F1平均值为89.9%,识别速率约为20 f/s。本文算法在刺梨果实的识别上取得了理想的识别效果。
The traditional target detection method relies on manual to extract features,which has the defects of low detection efficiency,poor robustness and poor real-time performance.In this paper,the YOLO v3 network with residual module is used according to the growth characteristics of the Rosa roxbunghii fruit in the natural environment.The training of the Rosa roxbunghii fruit recognition model is carried out by extracting the feature maps of different convolutional layers,and then extracting the deep feature maps and superimposing them with the shallow feature maps to extract deeper feature information in the images.Through the adjustment and improvement of the relevant parameters of the network,and testing 70 images of Rosa roxbunghii that did not participate in the model training,the experiment shows that the algorithm can effectively identify the 11 types of Rosa roxbunghii fruits in the natural environment.The average accuracy rate was 88.5%,the average recall rate was 91.5%,and the average value of F1 was 89.9%,and the recognition rate is about 20 f/s.The algorithm has achieved an ideal recognition effect on the identification of the Rosa roxbunghii fruit.
作者
闫建伟
赵源
张乐伟
张富贵
Yan Jianwei;Zhao Yuan;Zhang Lewei;Zhang Fugui(College of Mechanical Engineering of Guizhou University/Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education,Guiyang,550025,China;Research Center of Rosa roxbunghii Engineering Technology of National Forestry and Prairie Bureau,Guiyang,550025,China)
出处
《中国农机化学报》
北大核心
2020年第10期191-196,共6页
Journal of Chinese Agricultural Mechanization
基金
贵州省普通高等学校工程研究中心建设项目(黔教合KY字[2017]015)
贵州省科技计划项目(黔科合重大专项字[2019]3014—3、黔科合成果[2019]4292号、黔科合平台人才[2019]5616号)。
