Identification of tomato diseases effectively using an attention-based EfficientNet model

Computer vision enables machines to interpret and understand visual data. Deep convolutional neural networks (DCNNs), specifically designed to classify images, have revolutionized the field of computer vision. In spite of their success, these models often struggle to precisely localize critical regions of interest and encounter difficulties in distinguishing between images with highly similar characteristics, leading to classification errors. Agriculture plays a pivotal role in global sustainability. Therefore, accurate identification of plant diseases is crucial to mitigate substantial economic losses. Tomato is an economically important crop, and tomato diseases often exhibit symptoms that could be confusingly similar, making accurate diagnosis challenging. To address localization and similarity issues in tomato diseases, this study proposes a multi-head attention-based EfficientNetB0 model. The proposed model was trained and tested on field-condition images of tomato diseases. The tomato dataset consists of nine tomato diseases, including healthy images, and has a total of 8789 images. The proposed model achieved 97.4% test accuracy, which is a slight improvement compared to the baseline EfficientNetB0 model. Also, the proposed model exhibited the lowest rate of misclassifications. Furthermore, the Score-CAM technique was employed to evaluate the effectiveness of the proposed model, and explanations generated by the Score-CAM demonstrated the significant performance of the model in localizing disease symptoms.