基于深度学习的柠檬尺寸自动检测  

Automatic lemon size detection based on deep learning

作　　者：李帅 安维胜[1] 程卫东 钱华政 Li Shuai;An Weisheng;Cheng Weidong;Qian Huazheng(College of Mechanical Engineering,Southwest Jiaotong University,Sichuan Chengdu,610031,China;Tongwei Co.,Ltd.,Sichuan Chengdu,610014,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031 [2]通威股份有限公司,四川成都610014

出　　处：《机械设计与制造工程》2024年第8期86-90,共5页Machine Design and Manufacturing Engineering

基　　金：中国博士后科学基金面上项目(2021M702711);四川省科技技术项目(2019YFSY0003)。

摘　　要：为实现柠檬的自动分级,提出了一种基于改进RCF网络的阈值-面积算法自动检测柠檬尺寸的方法。首先在主干网络中加入注意力机制模块,在每个Stage侧级输出和最终输出时加入残差模块来改进RCF网络,以获得柠檬边缘更细致的检测图像;然后将边缘图像经过阈值处理后进行轮廓检测,运用基于OpenCV的阈值-面积算法提取柠檬轮廓,获得柠檬尺寸。实验结果表明:改进的RCF网络将全局最佳和最佳图像比例单图最佳指标分别提高到0.816和0.836;在尺寸检测实验中,阈值-面积算法检测的柠檬尺寸与实际尺寸相比误差较小。To achieve the automatic classification of lemons,a method based on the improved RCF network is proposed for the automatic detection of lemon sizes.Firstly,the attention mechanism module is incorporated into the network backbone,and the residual module is added to the side output and final output of each Stage to enhance the RCF network,thereby obtains more detailed detection images of lemon edges.Subsequently,the edge image is processed by thresholding for contour detection,and the lemon contour is extracted using the threshold-area algorithm based on OpenCV to acquire the lemon size.The experimental results demonstrate that the global optimal index and the optimal image ratio of a single image are increased to 0.816 and 0.836 respectively by the improved RCF network.In the size detection experiment,the error of the lemon size detected by the threshold-area algorithm is smaller than that of the actual size.

关 键 词：柠檬分级 边缘检测 注意力机制 OPENCV 阈值检测 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]