Analysis and Detection of Maize Seed Quality Using Lightweight Models

Abstract

Agriculture is one of the fastest-growing industries in India, but providing quality food to its large population remains a serious concern. Seed quality is of central importance to crop productivity, as low-quality seeds tend to result in decreased yields, higher production costs, and food insecurity. To address this issue, solve, this study proposes a real-time lightweight maize seed quality inspection system based on state-of-the-art You Only Look Once (YOLO) object detection models, including YOLOv5s6, YOLOv9, YOLOv11, and an Improved YOLOv8 (I-YOLOv8). The system distinguishes maize seeds into four classes— Pure, Discolored, Silk Cut, and Broken—based on a custom-labeled dataset of 2,128 images. By capitalizing on augmented small-object detection and better localization, the models were tested on important performance measures such as precision, recall, and mean Average Precision (mAP). The outcome of this comparison indicates that YOLOv5s offers the best trade-off between precision (72.0%) and mAP (76.7%), whereas I-YOLOv8 and YOLOv9 excel in terms of detection speed. This study highlights the accuracyversus. efficiency trade-offs and presents timely insights for selecting the most appropriate YOLO model to ensure real-time maize seed quality detection. The envisioned system helps enhance seed quality evaluation, thereby ensuring sustainable agricultural practices and enhancing food security.