基于深度学习的水果识别系统设计  

Design of Fruit Recognition System Based on Deep Learning

作　　者：李赟[1] 刘思雨 朱川[1] 常兴治[1] Li Yun;Liu Siyu;Zhu Chuan;Chang Xingzhi(Changzhou College of Information Technology,Changzhou 213000,China)

机构地区：[1]常州信息职业技术学院软件与大数据学院,江苏常州213000

出　　处：《农机化研究》2023年第10期187-191,共5页Journal of Agricultural Mechanization Research

基　　金：江苏省高等学校自然科学研究面上项目(19KJB520023);江苏省发改委省级工程研究中心建设项目(2019);常州信息职业技术学院校级项目(CXZK202011Z);江苏省高等教育教学改革研究项目(2019JSJG467);常州大学高职教育研究院项目(CDGZ2019045)。

摘　　要：为解决水果分拣难题,设计了一种基于深度学习的水果识别系统。首先,利用网络爬取工具对水果图片进行抓取、清洗和分类,完成数据集制作;然后,基于SSD(Single Shot Mutibox Detector)目标检测算法开展数据集的学习训练,生成权值文件;最后,将权值文件载入预训练模型并对水果图片进行预测,完成基于PyQt5框架的前段界面设计,实现水果边界框和分类置信度的图形展示。试验结果表明:系统可对视频文件进行快速、准确的水果图像识别,视频识别实时帧率稳定在22fps/s左右,识别准确率大于85%,可为后续水果识别、分拣等自动化作业提供技术支撑。In order to solve the problem of fruit sorting,a fruit recognition system based on deep learning was designed.Firstly,a network crawling tool was used to capture,clean and classify the fruit pictures,and completed the production of the data set;Then,the data set was trained based on SSD(single shot mutibox detector)target detection algorithm to generate the weight file;Finally,the weight file was loaded into the pre-training model to predict the fruit pictures,and the front interface was designed based on pyqt5 framework to realize the graphical display of fruit boundary box and classification confidence.The experimental results showed that the system can recognize fruit images quickly and accurately.The real-time frame rate of video recognition was stable at about 22 FPS/s,and the recognition accuracy was more than 85%,which could provide technical support for subsequent automatic operations such as fruit identification and sorting.

关 键 词：水果识别 深度学习 目标检测 SSD算法 

分 类 号：S24[农业科学—农业电气化与自动化] TP391.41[农业科学—农业工程]