Performance evaluation of hyper-parameter tuning automation in YOLOV8 and YOLO-NAS for corn leaf disease detection

Abstract

Corn cultivation was crucial in Southeast Asia, significantly contributing to regional food security and economies. However, leaf diseases posed a significant threat, causing substantial losses in production and quality. This research utilized artificial intelligence (AI) technology to address this issue by automating the hyper-parameter tuning process in YOLO (You Only Look Once) object detection models for early corn leaf disease detection. High-resolution images of corn leaves were captured and preprocessed for consistency. The preprocessing stage involved creating new dataset folders for images and labels, resizing images while preserving their aspect ratio, and rotating them if necessary. The images, containing 11,596 labeled instances, were analyzed using YOLOv8 and YOLO-NAS models. Each image's detected disease regions were converted into YOLO-format text files with x, y, width, and height coordinates, describing the presence and severity of infections. The models' performances were evaluated using precision, recall, mAP50, and mAP50-95 metrics. YOLOv8m achieved a mAP50 of 98.5% and mAP50-95 of 67.8%, while YOLO-NAS-L demonstrated superior detection capabilities with a mAP50 of 70.3% and mAP50-95 of 38.9%. This automated system facilitated early disease identification and enabled prompt preventive measures, thereby enhancing crop yields and mitigating losses. The findings highlighted the potential of advanced AI-driven detection systems in revolutionizing crop management and supporting global food security.