基于快速卷积神经网络的果园果实检测试验研究  

Experimental research on orchard fruit detection based on fast convolutional neural network

作　　者：张磊[1] 姜军生[2] 李昕昱 宋健[2] 解福祥[2] Zhang Lei;Jiang Junsheng;Li Xinyu;Song Jian;Xie Fuxiang(College of Mechanical and Electronic Engineering,Shandong University of Science and Technology,Qingdao,266590,China;College of Mechanical and Vehicle Engineering,Weifang University,Wei fang,261061,China;Shandong Institute of Industrial Technicians,Weifang,261061,China)

机构地区：[1]山东科技大学机械电子工程学院,山东青岛266590 [2]潍坊学院机电与车辆工程学院,山东潍坊261061 [3]山东工业技师学院,山东潍坊261061

出　　处：《中国农机化学报》2020年第10期183-190,210,共9页Journal of Chinese Agricultural Mechanization

基　　金：国家自然科学基金项目(51505337);山东省重点研发计划项目(2019GNC106144);山东省农机装备研发创新技术项目(2018YF005-05);山东省高等学校科技计划项目(J17KA150)。

摘　　要：近年来,基于数字图像处理和机器学习算法的果实自动识别检测研究已经越来越成熟。针对传统检测方法检测过程中难以满足实时性要求的缺点,采用了基于Faster-RCNN的果实快速检测模型。模型由卷积神经网络(CNN)和区域提议网络(RPN)组成,首先由CNN进行卷积和池化操作提取特征,然后由RPN选取候选区域,通过网络全连接层参数共享,由目标识别分类器和边界框预测回归器得到多个可能包含目标的预测框,最后通过非极大值抑制挑选出精度最高的预测框完成目标检测。分别对桃子、苹果和橙子的三种果实进行检测,采用迁移学习方法,使用已经预训练好的两种深度神经网络模型ZFnet和VGG16,通过数据集的训练对Dropout及候选区域数量进行参数调整完成网络调优。检测并分析果实不同布局形态下模型的检测效果。试验结果表明,当Dropout取值为0.5或0.6,候选区域数量为300时网络模型最佳,ZFnet网络中,苹果平均精确度为92.70%,桃子为90.00%,而橙子为89.72%。VGG16网络中,苹果平均精度为94.17%,桃子为91.46%,橙子为90.22%。且ZFnet和VGG16的图像处理速度分别达到17 fps和7 fps,能够达到果实实时检测的目的。In recent years,research on automatic fruit recognition and detection based on digital image processing and machine learning algorithms has become more and more mature.Aiming at the shortcomings of the traditional detection method that it is difficult to meet the real-time requirements,a fast fruit detection model based on Faster-RCNN is adopted.The model is composed of Convolutional Neural Network(CNN)and Region Proposal Network(RPN).First,CNN performs convolution and pooling operations to extract features,then RPN selects candidate regions,shares the parameters through the fully connected layer of the network,and classifies by target recognition The predictor and the bounding box prediction regress or obtain multiple prediction frames that may contain the target,and finally select the prediction frame with the highest accuracy by non-maximum suppression to complete the target detection.The three fruits of peaches,apples and oranges were tested respectively,using the transfer learning method,using two pre-trained deep neural network models ZFnet and VGG16,and adjusting the parameters of Dropout and the number of candidate regions through the training of the data set.Network tuning.Detect and analyze the detection effect of the model under different layout patterns of fruits.The test results show that when the value of Dropout is 0.5 or 0.6,and the number of candidate regions is 300,the result has the best accuracy and recall rate.In ZFnet network,the average accuracy of apple is 92.70%,that of peach is 90.00%,and that of orange is 89.72%.In VGG16 network,the average precision of apple,peach and orange is 94.17%,91.46%and 90.22%.The image processing speed of ZFnet and VGG16 reaches 17 fps and 7 fps respectively,which can achieve real-time detection of fruits purpose.

关 键 词：果实检测 深度学习 农业机器人 神经网络 迁移学习 

分 类 号：S24[农业科学—农业电气化与自动化] TP24[农业科学—农业工程]