Surface Defect Detection of “Yuluxiang” Pear Using Convolutional Neural Network with Class-Balance Loss

With increasing consumer expectations for the quality and safety of agricultural products, intelligent detection gradation have considerable significance in production. The surface defect is an important indicator quality, but classified mainly using inefficient manual identification “Yuluxiang” pears. Because uncertainty high difficulty image acquisition agriculture, data imbalance between categories a common problem. For resolution these problems, class balance (CB) was used to re-weight sigmoid cross-entropy loss (SGM-CE), softmax (SM-CE), focal (FL) functions this study. CB-SGM-CE, CB-SM-CE, CB-FL were construct GoogLeNet network as convolutional neural (CNN) generalized feature extractor, transfer learning combined build models, respectively. results showed that better than SGM-CE, SM-CE, FL, achieved best (F1 score 0.993–1.000) 3 CB functions. Then, VGG 16, AlexNet, SqueezeNet, MobileNet V2 networks based on learning, Machine (ML) CNN classification models Compared with ML other 4 CB-FL-GoogLeNet model (accuracy 99.78%). A system developed. testing accuracy 95.28% system. This study realizes pear unbalanced dataset, provides method agriculture.