Classification of papaya fruit based on maturity using machine learning and transfer learning approach
Subject Areas : Sustainable production technologies
mohammad ghorbani
1
,
Mostafa Ghazizadeh-Ahsaee
2
,
Kazem Jafari Naeimi
3
1 - Biosystems Engineering Dept. Faculty of Agriculture, Shahid Bahonar University of Kerman, Kerman, Iran.
2 - Computer engineering department, Shahid Bahonar University of Kerman, Kerman, Iran
3 - Associate Professor، Department of Biosystems Engineering، Faculty of Agriculture، Shahid Bahonar University of Kerman، Kerman، Iran.
Keywords: Feature descriptors , Grading, Machine learning, Maturity, Transfer learning,
Abstract :
Grading and packing fruits based on visual inspections can be time-consuming, destructive, and unreliable. The objective of the conducted research is to provide an intelligent, fast and reliable classification method to detect the maturity of papaya fruit in three levels: immature, partially mature and mature. The total number of images used in this article is 300 images, 100 images have been collected for each level. In this paper, the use of two approaches, machine learning and transfer learning, is proposed to classify papaya fruit maturity status. The machine learning approach includes the use of three feature descriptors and three different classifiers, which are: local binary pattern (LBP), Gray Level Cooccurrence Matrix (GLCM), histogram of oriented gradients (HOG), k-nearest neighbor (KNN) classification algorithm, support vector machine(SVM) and Naïve Bayes classification algorithm. Transfer learning methods include the use of six pre-trained deep learning models Alexnet, Googlenet, Resnet101, Resnet50, Resnet18, VGG19. KNN classifier using HOG feature descriptor has achieved 95.4% accuracy and 3:52 seconds training time. The classifier based on transfer learning approach VGG19 was able to record better performance among other deep learning networks by obtaining 100% accuracy and training time of 10:42 seconds. Two classification methods using machine learning and transfer learning methods have been able to obtain accuracy of 95.4% and 100%, respectively, which are 0.7% and 6% more than the existing proposed methods.
