A fruit detection algorithm based on R-FCN in natural scene
作者单位：School of Information and Control Engineering Shenyang Jianzhu University
会议名称：《第32届中国控制与决策会议》
会议届次：32
主办单位：IEEE Control Systems Society (CSS);Northeastern University;State Key Laboratory of Synthetical Automation for Process Industries;Technical Committee on Control Theory, Chinese Association of Automation
会议日期：2020年
学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 0828[工学-农业工程] 08[工学] 080203[工学-机械设计及理论] 0835[工学-软件工程] 0802[工学-机械工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）]
基 金：Scientific research project in Liaoning Province Natural Science Foundation of China, (201602616) Scientific research project of Jiangsu Key Laboratory, (JSKLE201707)
关 键 词：Deep learning Region Proposal Network Fully Convolutional Network Target recognition and location
摘 要：Aiming at the problem of poor precision and low efficiency of the vision system when agricultural robot picks fruit, this paper effectively fuses deep learning with machine vision, and proposes a new algorithm for fruit recognition and location, R-FCN, a deep learning algorithm combining regional suggestion network（RPN） and full convolutional neural network（FCN）. The proposed algorithm uses FCN to convolve the input image to achieve feature extraction at the pixel level. Among them, the fusion of residual network can make the deep network have more abundant feature information for fruit recognition, and deconvolution can realize the visualization of detection results. Using RPN generates multiple candidate boxes on the feature map after convolution operation, this can effectively separate the image foreground and background. Training was conducted on the public COCO data set, and testing experiments were conducted on the different states of three different fruits. The experimental results show that the algorithm in this paper improves the detection accuracy by 0.71% and 0.33% respectively compared with the previous algorithm in apple and orange recognition. It can also have a recognition accuracy of 82.30% for banana, which is a large-scale fruit. For different input images, it can realize the visualization of fruit recognition and location, reduce the influence of branches and leaves occlusion, enhance the robustness of the system, and improve the efficiency of picking.