Application of Transfer Learning for Fruits and Vegetable Quality Assessment

In this paper, we utilize the concept of transfer learning in fruits and vegetable quality assessment. The transfer learning concept applies the idea of reuse the pre-trained Convolutional Neural Network to solve a new problem without the need for large-scale datasets for training. Eight pre-trained deep learning models namely AlexNet, GoogleNet, ResNet18, ResNet50, ResNet101, Vgg16, Vgg19, and NasNetMobile are fine-tuned accordingly to evaluate the quality of fruits and vegetable. To evaluate the training and validation performance of each fine-tuned model, we collect a dataset consists of images from 12 fruits and vegetable samples. The dataset builds over five weeks. For every week 70 images collected therefore the total number of images over five weeks is 350 and the total number of images in the dataset is (12∗350) 4200 images. The overall number of classes in the dataset is (12∗5) 60 classes. The evaluation of the models was conducted based on this dataset and also based on an augmented version. The model's outcome shows that the Vgg19 model achieved the highest validation accuracy over the original dataset with 91.50% accuracy and the ResNet18 model scored the highest validation accuracy based on the augmented dataset with 91.37% accuracy.