Application of transfer learning and image augmentation technology for tomato pest identification

Abstract

Traditional methods of identifying pests are time consuming and limited. To solve these problems, a set of reliable prediction models should be established to correctly identify tomato pests. This study established a tomato pest image dataset of common tomato pests. Deep learning (DL) convolutional neural network (CNN) models were applied to classify eight categories of tomato pests. Transfer learning was used to reduce training time. The DL models were used to extract features, and the extracted pest features were combined with three machine learning (ML) classifiers, including discriminant analysis (DA), support vector machine, and k-nearest neighbor method (KNN). Hyper-parameters were automatically optimized through Bayesian optimization. After image augmentation, VGG16 model exhibited the highest performance than the other models, with an accuracy of 94.95 %. Regarding the CNN + ML models, the ResNet50 with discriminant analysis model achieved 97.12 % classification accuracy. Combining the advantages of DL and ML could not only simplify the feature extraction process and reduce the training time, but also provide experts and farmers with effective and immediate assistance in identifying pests, which could help reduce economic and crop yield loss.