An In-Depth Study on Object Detection Techniques: A Case Study on Wheat Leaf Disease Detection
Abstract
In India, where agriculture contributes nearly 18% to the GDP, the economic and food security impact of crop failures is substantial. As the second-largest global producer of wheat, contributing to an annual production of around 109.6 million tonnes, India encounters substantial agricultural losses ranging from 15% to 25% attributed to issues such as weeds, diseases, and insects. With 58% of the population dependent on agriculture for income, early plant disease detection is crucial for crop quality improvement and economic stability. Deep Learning approaches, known for swift disease identification with minimal human resources, make a transformative shift in detection methodologies. This research addresses wheat leaf disease detection using fine-tuned object detection models. Utilizing a publicly available dataset featuring high-resolution images of three wheat leaf diseases-septoria, mildew, and brown rust-the Detectron2 (Faster R-CNN R50 FPN) and YOLOv8 models were fine-tuned and compared. Results revealed that YOLOv8 achieved slightly superior performance with 0.314 [email protected][IoU], making it the preferred model for our wheat leaf disease detection problem.
Keywords
Wheat leaf diseases, object detection, YOLOv8, Faster R-CNN, Fine-Tuning, Transfer Learning..