Plant disease detection with generative adversarial networks

Generative Adversarial Networks (GANs) have emerged as a promising approach for enhancing image detection and facilitating image classification. Deep learning models, which exhibit high classification accuracy, have proven advantageous over conventional approaches for Plant disease detection (PDD). However, since deep learning requires many training images, the generation of synthetic images resembling authentic images using GANs has been applied to datasets with a limited training size. Previous studies in computer vision have demonstrated that applying pre-trained classifiers to the generated synthetic images of plant diseases can increase accuracy by extracting useful features. To empirically validate the effectiveness of GANs on the performance of binary and multi-class PDD, we train GANs on diverse plant species and disease symptoms, enabling the classification of different plant diseases. To achieve this, we trained two GAN models, namely Deep convolutional GAN (DCGAN) and alpha beta GAN (αβGAN), on different groups and numbers of plant species and disease classes to generate synthetic im-ages. The generated synthetic images were used for PDD using the Visual Geometry Group (VGG16) model. Our analysis revealed that both DCGAN and αβGAN generated very realistic plant images. Further, we found that Images generated by DCGAN and αβGAN led to different predictions for. the same disease class, highlighting variations on how the two GANs represented features from the same plant-disease combination. Lastly, there were no significant differences between the performances of the model trained on the default dataset and the model trained on the large, modified datasets that included the synthetic images. Consequently, further research incorporating more generated images and fine-tuning GAN is required to address multi-class problems in PDD.