Classification of durian maturity using a convolutional neural network

Abstract

The maturity of a durian fruit is one of the important factors that are considered when assessing the appropriate quality of fruit for export worldwide. In order to verify the maturity, a high level of professional expertise and long-term experience are required. The best current method for classifying images is a convolutional neural network (CNN), a branch of deep learning. In this study, a CNN method is proposed to classify and predict the stages of maturity of durian fruit. In this experiment, durians were separated into five classes based on the harvesting period: (i) 96 days; (ii) 103 days; (iii) 110 days; (iv) 117 days; and (v) 124 days. Following this, the durian fruits were analyzed to define the dry weight and to allow the stages of maturity to be inspected. The results showed that the average dry weights were 11.31%, 17.76%, 26.65%, 38.94% and 42.13%, respectively. Mature durians have a dry weight of 32% or more, meaning that durians harvested at 117 and 124 days can be classified as mature, while the rest are classified as immature. Data were then collected for all five classes of durian, with 300 images per class, giving a total of 1,500 images. This study mainly focused on comparing three CNN architectures: LeNet-5, AlexNet, and DuNet-12 (our proposed CNN architecture). We also compared two activation functions, tanh and ReLU, and four optimization algorithms: AdaGrad, AdaDelta, RMSProp and Adam. The DuNet-12 architecture with the ReLU activation function and Adam optimizer was the most effective method over a training period of 350 epochs, and yielded a testing accuracy of 98.96%. It had the highest prediction accuracy of 100%. However, the results also demonstrated the efficiency of the proposed durian classification method, and the findings of our experiment could be used to develop equipment for the durian export industry in the future.