Identification of an Efficient Deep Learning Architecture for Tomato Disease Classification Using Leaf Images

Computer vision technology plays a vital role in studies on plant pathology. The study of plant disease using image processing refers to the study of visually observable patterns on the plants. Recently, various image processing and pattern classification techniques are used to develop digital vision systems that can identify and classify the visual symptoms of plant diseases. As there are so many algorithms for the identification of plant diseases through leaf image classification, it is very critical to know what algorithms provide high accuracy and what type of diseases can be identified by the algorithm. The main objective of this study was to present accurate deep learning architectures that were more efficient in detecting tomato diseases which would eliminate human error in the identification via naked-eye observation. A base model was built with Convolutional Neural Network (CNN) from scratch using 22930 images of tomato plant leaves and compared with architectures like VGG16, MobileNet, and Inceptionv3 by fine-tuning them. The base CNN model has shown a training accuracy of 90% while other models VGG16, MobileNet, and Inceptionv3 have shown 89%, 91%, 87% accuracies, respectively. The computation complexity of the VGG16 model was higher than the other methods because the number of parameters defined in the model was high. Nevertheless, the accuracy of MobileNet was the highest compared with others. It is perfect for the study as it is lightweight, faster and can be easily run on mobile devices. The significant difference between the proposed CNN architecture and the others is that, the proposed CNN is slightly shallower and can be trained on the same dataset much more quickly. This study will help new researchers to conduct further improvements in the context of plant disease classification without any difficulty.