Deep transfer learning classification of apple fruit diseases

Abstract

This paper applies deep convolution neural networks (DCNN) to apple fruit disease classification. Twelve DCNN methods (SqueezeNet, GoogleNet, InceptionV3, DenseNet201, ReaNet50, ResNet101, Xception, InceptionResnetV2, EfficientnetB0, AlexNet, VGG16, and VGG19) have been used. These methods have been trained to classify apples into four categories: normal, blotch, rot, and scab. A dataset of 5179 images, including 3472 for normal, 171 for blotch, 1166 for rot, and 370 for scab, has been used. A practical test on 120 images (30 for each category) has been applied. Seven of these DCNNs—InceptionV3, DenseNet201, ResNet101, ResNet50, GoogleNet, AlexNet, and VGG16—have the best accuracy. InceptionV3 is the highest. It has achieved an accuracy of 100% for all categories. The used dataset is unbalanced and small. So, it's necessary to use data augmentation to overcome any overfitting that may cause. After applying data augmentation, the dataset is balanced and contains 13888 images (3472 for each category). The seven DCNNs are retrained by the balanced dataset and retested by the same 120 images. All DCNN's accuracy has enhanced except InceptionV3, which has decreased. On the other hand, RasNet101 has achieved an accuracy of 100% for all categories. Therefore, ResNet101 has been recommended for apple fruit disease classification.