Classification of Durian Variety for Export Using Convolutional Neural Network

Abstract

This paper proposed the image processing technique using convolutional neural network (CNN) to increase the durian classification performance in accordance with the export standard. In this experiment, durians were sorted into 12 classes from 4 varieties and 3 quality class: Monthong Extra Class, Monthong Class I, Monthong Class II, Chanee Extra Class, Chanee Class I, Chanee Class II, Kanyao Extra Class, Kanyao Class I, Kanyao Class II, Kradumthong Extra Class, Kradumthong Class I, and Kradumthong Class II. The main focus was to compare LeNet-5 and simple architectures of convolution neural network. The result showed that the resizing image to 500x500 pixels using LeNet-5, Rectified Linear Units (ReLU), and 30 epochs had the highest training accuracy of 99.86%. While, the architecture of 4 convolution layers using the number of epochs for 50 epochs had the highest test accuracy of 95.36%. Then again, the architecture of 5 convolution layers using 50 epochs had the highest accuracy in prediction of 97.89%. The obtained results showed that the employed a simple architecture can accurately classify durian images and make predictions as accurate as 97.89%. The results show that using a simple architecture of convolution neural network. can accurately classify 500x500 pixel durian images and make the most accurate predictions.