基于改进残差网络的红枣缺陷检测分类方法研究  

Research on defect detection and classification of jujube based on improved residual network

作　　者：文怀兴[1] 王俊杰 韩昉[1] WEN Huai-xing;WANG Jun-jie;HAN Fang(College of Mechanical and Klectrical Kngineering,Shaanxi University of Science&Technology,Xi'an,Shaanxi 710021,China)

机构地区：[1]陕西科技大学机电工程学院

出　　处：《食品与机械》2020年第1期161-165,共5页Food and Machinery

基　　金：陕西省重点研发计划项目(编号：2019GY-024)

摘　　要：提出了一种基于深度残差网络对红枣表面缺陷以及纹理识别的分类算法,将红枣RGB彩色图的G分量图进行预处理后得到的特征图作为网络的输入,采用残差学习的方式扩大神经网络的学习深度,并将残差神经网络的激活函数Relu替换为SELU,对损失函数softmax loss用center loss进行替换,训练时引入Dropout层降低网络过拟合风险,解决了随着学习深度加深网络中梯度弥散和爆炸的现象。研究结果表明:该分类方法准确率达96.11%,检测效率约为120个/min。In this paper,an algorithm based on deep residual network is proposed to recognize and classify the surface defects and texture of jujube.The algorithm adopted jujube's G channel of RGB color figure then to get the characteristics of the figure as network input,using residual learning way to expand the depth of the neural network learning,and residual error of the neural network activation function Relu replaced with SELU,the loss function softmax loss with center loss to replace,Dropout layer was developed for the training,reduce the risk of network through fitting,solved with deepening study depth gradient dispersion and explosion phenomenon in the network.The results showed that the classification accuracy reached 96.11% and the detection efficiency was about 120 min^ 1.

关 键 词：红枣 残差网络 缺陷检测 激活函数 损失函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] S665.1[自动化与计算机技术—控制科学与工程] TP391.41[农业科学—果树学]