Durian Ripeness Classification Using Deep Transfer Learning

This study aims to investigate the application of deep transfer learning to the ripe-ness levels of the Monthong durian classification. The objective is to compare the effectiveness of different classification models, including Inception-v2, Inception-v3, VGGNet16, VGGNet19, ResNet-50, ResNet-101, ResNet-152, MobileNet-v2, and MobileNet-v3. The dataset for this research consists of 1000 images of the Monthong durian, divided into 4 levels: over-ripe, semi-ripe, unripe, and ripe. Hyperparameters were specified to enhance the accuracy of classification of the ripeness levels using pretrained and modified models. The deep transfer learning models and the top 3 performing models were MobileNet-v2 with an accuracy of 95.50%, VGG16 with an accuracy of 94.50%, and VGG19 with an accuracy of 94.50%. The deep transfer learning model with the lowest performance was Inception-v2, achieving an accuracy of 81.50%, particularly in identifying the complex characteristics of the Monthong durian. We anticipate that the results of this study will not only contribute to advancements in durian ripe-ness classification but also provide valuable insights for stakeholders in the durian industry. This will guide cultivation practices and marketing strategies with more comprehensive data while examining how this research contributes to the classification of agricultural practices and the application of machine learning in a broader range of industrial sectors.