Plant Disease Detection Using a Hybrid Approach Based on Vision Neural Network Transformers

Abstract

Plant disease detection is one of the main challenges in the agricultural sector, acknowledged as a significant contributor to crop losses, negatively impacting food production and the global economy. In response to this issue, novel agricultural solutions are emerging, leveraging the synergy of deep learning and computer vision technique for early disease detection. Recently, researchers have embraced vision Transformers for plant disease identification. While this approach has shown promising results, it also presents challenges, such as high computational costs and low inductive bias for locality. In response to these challenges, we suggest a hybrid model with fewer trainable parameters, combining the power of vision Transformers with the capabilities of convolutional layers to extract relevant local features, thereby enhancing the performance of ViT. Additionally, we modify the ViT architecture to minimize the number of trainable parameters. We performed experiments on a public dataset of potatoes to assess the effectiveness of our proposed hybrid model. According to the experimental results obtained, the proposed model achieved a test accuracy of 98.27%, surpassing the original ViT and two leading CNN architectures, namely VGG16 and ResNet50. Our model has fewer parameters than the original ViT, with a reduction rate of 49%. This reduction helps to minimize the computational and memory costs of our model. Additionally, due to this parameter reduction, our proposed model requires half of training time compared to the original ViT. All these findings make our proposed model more adaptable for smart agriculture applications, particularly in real-time detection of plant diseases.