Tomato Disease Recognition Using a Compact Convolutional Neural Network

Abstract

The detection of diseases in tomatoes in advance and early intervention and treatment increase the production amount, efficiency, and quality, which will satisfy the consumer with a more affordable shelf price. Thus, the efforts of farmers waiting for harvests throughout the season are not wasted. In this study, a compact convolutional neural network (CNN) is proposed for a disease identification task in which the network comprises only six layers, which is why it is computationally inexpensive in terms of the parameters employed in the network. This network was trained using PlantVillage's tomato crop dataset, which consisted of 10 classes (nine diseases and one healthy). The proposed network was first compared with the well-known pre-trained ImageNet deep networks using a transfer learning approach. The results show that the proposed network performs better than pre-trained knowledge transferred deep network models, and that there is no need to constitute very large, complicated network architectures to achieve superior tomato disease identification performance. Furthermore, data augmentation techniques are employed during network training to improve the performance of the proposed network. The proposed network achieved an accuracy of the F-1 score, Matthews correlation coefficient, true positive rate, and true negative rate of 99.70%, 98.49%, 98.31%, 98.49%, and 99.81%, respectively, using 9,077 unseen test images. Our results are better than or similar to those of state-of-the-art deep neural network approaches that use the PlantVillage database and the proposed method employs the cheapest architecture.