复杂野外环境下油茶果快速鲁棒检测算法  

Fast and robust detection algorithm for camellia oleifera fruit in complex field environment

作　　者：周浩[1] 唐昀超 邹湘军[1] 王红军[1] 陈明猷 黄钊丰 ZHOU Hao;TANG Yunchao;ZOU Xiangjun;WANG Hongjun;CHEN Mingyou;HUANG Zhaofeng(College of Engineering,South China Agricultural University,Guangzhou 510630,China;College of Urban and Rural Construction,Zhongkai University of Agriculture and Engineering,Guangzhou 510080,China)

机构地区：[1]华南农业大学工程学院,广东广州510630 [2]仲恺农业工程学院城乡建设学院,广东广州510080

出　　处：《现代电子技术》2022年第15期73-79,共7页Modern Electronics Technique

基　　金：广东省科技计划项目(2019A050510035);广东省农业厅科技项目(粤农农函[2019]1019号)。

摘　　要：为了提高移动采摘机器人在复杂野外环境下检测油茶果的速度和鲁棒性,在YOLOv4⁃tiny网络的基础上提出YOLO⁃Oleifera网络。首先将两个1×1和3×3的卷积核分别添加至YOLOv4⁃tiny网络的第2个和第3个CSPBlock模块之后,以有助于学习油茶果的特征信息和减少计算复杂度;接着使用K⁃means++先验框聚类算法代替YOLOv4⁃tiny网络使用的K⁃means先验框聚类算法,以获得满足油茶果尺寸的聚类结果。消融实验证明了网络改进的有效性。分别测试光照和阴影环境下的油茶果图像,实验表明YOLO⁃Oleifera网络在不同光照条件下检测油茶果具有鲁棒性。此外,对比实验表明被遮挡的油茶果因为语义信息的缺失而导致Precision和Recall降低。将YOLO⁃Oleifera网络的测试结果与YOLOv5⁃s、YOLOv3⁃tiny和YOLOv4⁃tiny网络进行比较,结果显示YOLO⁃Oleifera网络的AP最高,而且YOLO⁃Oleifera网络占用硬件资源最小。此外,YOLO⁃Oleifera网络检测图像平均花费31 ms,能够满足移动采摘机器人的实时检测需求。因此,提出的YOLO⁃Oleifera网络更加适合搭载在移动采摘机器人上进行检测任务。A YOLO⁃Oleifera network based on YOLOv4⁃tiny network is proposed to increase the detecting speed and improve the detecting robustness of mobile picking robot in the picking of camellia oleifera fruit in complex field environment.In this study,1×1 and 3×3 convolution kernels are added to the 2nd and 3rd CSPBlock modules of YOLOv4⁃tiny network respectively,so as to facilitate the learning of the feature information of camellia oleifera fruit and reduce the computational complexity.And then,K⁃means++prior box clustering algorithm is used to replace the K⁃means prior box clustering algorithm used in YOLOv4⁃tiny network,so as to obtain the clustering results that meet the size of camellia oleifera fruit.The ablation experiment was performed to prove the effectiveness of the network improvement.The images of camellia oleifera fruit in sunlight and shading were tested.The results shows that,in different illumination conditions,YOLO⁃Oleifera network is robust in detecting camellia oleifera fruit.Furthermore,the contrast experiment shows that the detection precision and recall of occluded camellia oleifera fruit is decreased due to the loss of semantic information.The test results of the YOLO⁃Oleifera network are contrasted with those of the YOLOv5⁃s network,YOLOv3⁃tiny network and YOLOv4⁃tiny network.The results show that the YOLO⁃Oleifera network has the highest AP among them,and it occupies the least hardware resources.In addition,it takes an average of 31 ms to detect images by the YOLO⁃Oleifera network.Therefore,the YOLO⁃Oleifera network can meet the real⁃time detection needs and is more suitable to be mounted on the mobile picking robot for detection.

关 键 词：目标检测 YOLOv4⁃tiny网络 深度学习 卷积核 采摘机器人 K⁃means++ 鲁棒性 

分 类 号：TN711-34[电子电信—电路与系统] TP391.4[自动化与计算机技术—计算机应用技术]