Potato Leaf Disease Detection: Comparative Study of the Performance of CNN Architectures - VGG19, ResNet50 and DenseNet201

Rodriguez J.L.R.; Rodriguez J.M.R.; Sanchez J.M.P.; Moreno L.A.O.; Corro I.D.M.; Ramirez E.D.V. · ICSPIS 2023 - Proceedings of the 9th International Conference on Signal Processing and Intelligent Systems (2023) · DOI: 10.1109/icspis59665.2023.10402706

Resumen

In this study, the performance of the CNN architectures VGG19, ResNet50, and DenseNet201 is compared in the detection of diseases in potato leaves. Within the scope of this research, the Plant Village database (2152 images) was employed. To this database composed of healthy potato leaves, late blight-infected leaves, and early blight-infected leaves, data augmentation with Gaussian Blur was applied, resulting in a total of 2760 images. The CNN architectures were trained using transfer learning with different batch sizes (16, 32, and 64). Based on three parameters, it was determined that the CNN architecture with the best performance was DenseNet201 with a batch size of 16, achieving an accuracy of 99.64%, requiring a storage capacity of 79.9 MB, and a training time of 16 minutes and 16 seconds.

Instituciones: Universidad Nacional de Trujillo