Deep learning-based detection and counting of wheat seeds: Comparative benchmarking of YOLO models

Automated wheat-seed detection and counting play important roles in high-throughput plant phenotyping, seed characterization, and agricultural research. Conventional manual counting is labor-intensive, time-consuming, and susceptible to human error when processing large numbers of seed samples. Recent advances in deep learning and object detection provide opportunities to automate these tasks using conventional RGB images. This study focused on a deep learning-based framework for wheat-seed detection and detection-based counting using a custom red-green-blue (RGB) image dataset. A dataset comprising 832 RGB images containing 3,649 manually annotated wheat-seed instances was developed, with images representing one to ten detached wheat seeds per image. All seed instances were manually annotated using bounding boxes and formulated as a single-class object detection problem. Two lightweight object detection models, YOLOv8n and YOLO11n, were trained and evaluated under identical experimental conditions. The model performance was assessed using precision, recall, mean Average Precision at an Intersection over Union (IoU) threshold of 0.5 (mAP@0.5), mean Average Precision averaged across IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95), training loss curves, confidence-based performance curves, confusion matrices, and qualitative detection outputs. Both models achieved excellent detection performance on the custom wheat seed dataset. YOLOv8n achieved a precision of 0.9947, recall of 0.9963, mAP@0.5 of 0.9940, and mAP@0.5:0.95 of 0.5284. YOLO11n produced slightly higher performance, achieving a precision of 0.9988, recall of 0.9984, mAP@0.5 of 0.9950, and mAP@0.5:0.95 of 0.5385. Overall, the performance of the two models was similar at mAP@0.5, but at mAP@0.5:0.95 stricter criterion, YOLO11n consistently demonstrated the strongest overall performance. The findings show that lightweight YOLO models combined with RGB imaging provide an effective and easy way for automated localization and counting of wheat seeds. ​The framework provides a manually annotated RGB wheat-seed dataset and a reproducible benchmark to compare lightweight YOLO models for detection. This study provides a practical foundation for future research on automated seed phenotyping, with future work focusing on external validation using more diverse datasets, multi-class seed-quality assessment, and quantitative evaluation of counting performance.