基于PP-YOLO改进算法的脐橙果实实时检测  

Real-time Detection of Navel Orange Fruit Based on Improved PP-YOLO Algorithm

作　　者：章倩丽 李秋生[1,2] 胡俊勇 谢湘慧 ZHANG Qianli;LI Qiusheng;HU Junyong;XIE Xianghui(Research Center of Intelligent Control Engineering Technology,Gannan Normal University,Ganzhou Jiangxi 341000,China;School of Physics and Electronic Information,Gannan Normal University,Ganzhou Jiangxi 341000,China)

机构地区：[1]赣南师范大学智能控制工程技术研究中心,江西赣州341000 [2]赣南师范大学物理与电子信息学院,江西赣州341000

出　　处：《北京联合大学学报》2022年第4期58-66,共9页Journal of Beijing Union University

基　　金：国家自然科学基金资助项目(42061067);江西省教育厅科学技术研究项目(GJJ201408)。

摘　　要：深度学习已被广泛应用于智能采摘领域,消除不同环境场景对目标识别和检测产生的不利影响,对采摘机器人实现精准高效的工作至关重要。采用基于单阶段目标检测网络改进的PP-YOLO模型对树上成熟脐橙的识别进行研究,通过添加可变形卷积的主干网络ResNet提取特征,结合特征金字塔网络(FPN)进行特征融合,实现多尺度检测,并采用K-means聚类算法得到与目标脐橙适宜的anchor尺寸,减少训练时间及预测框置信度误差。实验结果表明:改进的PP-YOLO检测模型可完成晴天逆光、晴天顺光和阴天环境下的脐橙检测任务,检测准确率分别为90.81%、92.46%和94.31%,检测速度可达到72.30 fps、73.71 fps和74.90 fps,可以尝试在脐橙采摘机器人的研制中加以应用。Deep learning is widely used in the field of intelligent picking,and eliminating the adverse effects of different environmental scenes on target detection and recognition is crucial to the accurate and efficient work of picking robots.In this paper,the improved PP-YOLO model based on the one-stage target detection network is used to study the recognition of ripe navel oranges on trees.By adding the deformable ResNet backbone network to extract features of convolution and combining the feature pyramid network(FPN)for feature fusion,multi-scale detection is realized.In addition,the paper uses the K-means clustering algorithm to cluster the anchor size suitable for the target navel orange to reduce the training time and the confidence error of the prediction frame.The experimental results show that the improved PP-YOLO detection model can achieve navel orange detection tasks in sunny day backlight,sunny day smooth light and cloudy day conditions,with the detection accuracy of 90.81%,92.46%and 94.31%respectively,and the recognition speed of 72.30 fps,73.71 fps and 74.90 fps respectively.It can be applied in the development of navel orange picking robot.

关 键 词：脐橙 目标检测 深度学习 改进的PP-YOLO 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]