Transfer Learning-based Object Detection Models for Improved Diagnosis of Tomato Leaf Disease

Abstract

During their growth, tomatoes are impacted by several illnesses and pests. The most crucial step in effectively assisting vegetable farmers in increasing tomato yield is combating diseases by the precise and early identification of these problems. Utilizing deep learning for detecting objects saves time, effort and allows real-time judgment, thereby reducing the substantial losses in the yield. This study uses the Open-source dataset of Tomato Leaf Diseases under natural conditions and trains them with multiple YOLO models(YOLOv5, YOLOv7) for the characteristics of tomato diseases’ images. Transfer learning is subsequently applied to augment the adaptability of these YOLO models, while the integration of a YOLO-YOLO cascaded ensembling technique enhances multiscale feature detection. This approach significantly elevates the accuracy of disease detection, effectively pinpointing both the location and classification of tomato leaf diseases. The prominent YOLO v5-v7 ensemble transfer learning model, when evaluated, exhibits noteworthy performance metrics during training. It achieves a mean average precision (mAP) of 98.8% at a detection threshold of 0.5 and a mAP of 98.7% at a more stringent threshold of 0.95. Additionally, the model demonstrates high precision at 95.8% and a strong recall rate of 96.8%.