基于GoogLeNet深度迁移学习的苹果缺陷检测方法

Defect Detection Method of Apples Based on GoogLeNet Deep Transfer Learning

针对目前国内苹果分选大部分以人工操作的现状,提出利用GoogLeNet深度迁移模型对苹果缺陷进行检测。检测结果表明,本文方法对扩充后的1932个训练样本的识别准确率为100%,对235个测试样本的识别准确率为91.91%。为评估目前苹果缺陷检测常用算法的性能,将GoogLeNet与浅层卷积神经网络(AlexNet和改进型LeNet-5)及传统机器学习方法(K-NN、RF、SVM)进行了对比,结果表明,与苹果缺陷检测的常用算法相比,本文方法具有更好的泛化能力与鲁棒性。

Apple processing has been one of the most important aspects in the field of fruit and vegetable processing for a long time,and how to screen out the defects of apple with high precision and low cost has been one of the key research directions at home and abroad.In view of the current situation of fruit sorting which mainly completed by manual operation in China,the deep transfer model GoogLeNet based on deep convolutional neural network was used to detect the defects of apple,and the results showed that the accuracy rates of GoogLeNet could reach up to 100%and 91.91%based on 1932 expanded training samples and 235 testing samples,respectively.At the same time,through assessing the performance of common machine learning algorithms in the field of apple defects detection,the results of GoogLeNet were compared with the shallow convolutional neural network(AlexNet and the improved LeNet-5)and traditional machine learning algorithms(K-nearest neighbor,K-NN;random forest,RF;support vector machine,SVM)in order to further verify the superiority of GoogLeNet.The results indicated that deep convolutional neural network had better generalization ability and robustness when compared with other conventional algorithms in the field of apple defects detection,which supported its broad application prospects.