Enhanced classification performance through GauGAN‐based data augmentation for tomato leaf images

Abstract

AbstractThis study investigated a data augmentation method for plant disease classification and early diagnosis based on a generative adversarial neural network (GAN). In the development of classification models using deep learning, data imbalance is a primary factor that reduces classification performance. To address this issue, tomato disease images from the public dataset PlantVillage were used to evaluate the performance of the GauGAN algorithm. The images generated by the proposed GauGAN model were used to train a MobileNet‐based classification model and compared with methods trained with conventional data augmentation techniques and cut‐mix and mix‐up algorithms. The experimental results demonstrate that based on F1‐scores, GauGAN‐based data augmentation outperformed conventional methods by more than 10%. In addition, after the model was retrained on data collected in the field, it efficiently generated various disease images. The evaluation results from those images also revealed a data augmentation effect of about 10% compared with traditional augmentation techniques.