期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2017
卷号：95
期号：24
页码：6800
出版社：Journal of Theoretical and Applied
摘要：Deep Learning for plant leaf analysis has been recently studied in various works. In most cases, Transfer Learning has been utilized, where the weights of networks, which are stored in the pre-trained models, are fine-tuned to use in the considered task. In this paper, Convolutional Neural Networks (CNNs), are employed to classify tomato plant leaf images based on the visible effects of diseases. In addition to Transfer Learning as an effective approach, training a CNN from scratch using the Deep Residual Learning method, is experimented. To do that, an architecture of CNN is proposed and applied to a subset of the PlantVillage dataset, including tomato plant leaf images. The results indicate that the suggested architecture outperforms VGG models, pre-trained on the ImageNet dataset, in both accuracy and the time required for re-training, and it can be used with a regular PC without any extra hardware required. A common feature visualization and verification technique is also applied to the results and further discussions are made to imply the importance of background pixels surrounding the leaves.
