Real-Time Plant Disease Detection: A Comparative Study

Plant diseases account for over 30% of production loss in India. Early detection of these diseases is crucial to maintaining yield. However, manual surveillance is laborious, costly, time-consuming and requires domain knowledge. Computer vision offers a non-destructive and efficient solution to disease detection, with classical machine learning and deep learning (DL) algorithms. DL methods offer several advantages, especially in scenarios where there is a large amount of data to process. With automatic feature extraction, these techniques can efficiently analyze multi-dimensional inputs, reducing the time and effort required for processing. Consequently, their usage has gained significant popularity in identifying and diagnosing diseases in plants. In this paper, we conduct a comprehensive comparative study of 14 cutting-edge object detection algorithms, including default and modified versions of YOLOv7 and YOLOv8. Our study focuses on their performance in real-time plant disease detection. The study involved several stages, including pre-processing, fine-tuning using pre-trained weights and validation on two publicly available datasets, namely PlantDoc and Plants Final, comprising real-life images of plant leaves. In particular, the study compared the performance of default YOLO models with YOLO models that used default architecture after freezing the backbone weights during the As per Springer style, both city and country namesmust be present in the affiliations. Accordingly, we have inserted the country name in the affiliation. Please check and confirm if the inserted country name is correct. If not, please provide us with the correct country name fine-tuning process. The results show that the modified YOLO models achieved comparable mean Average Precision (mAP) values to the default models on the Plants Final dataset while reducing training time and GPU load by 9-25% and 19-47%, respectively, depending on the size of the backbone in different models.