Star-YOLO: A Lightweight Real-Time Wheat Grain Detection Model for Embedded Deployment

With the rapid advancement of precision agriculture, traditional object detection algorithms struggle with limited efficiency and accuracy in wheat grain detection and counting, while the need for real-time deployment of deep learning models on embedded devices becomes increasingly critical. To this end, this paper introduces Star-YOLO, a lightweight wheat grain detection model built upon YOLOv11n. The model employs StarNet to refine the C3k2 structure, reducing computational complexity without compromising detection accuracy, and integrates the MBConv module into the detection head to boost feature extraction while further minimizing computational load. A Shape-NWD loss function is designed, incorporating shape and scale information of target bounding boxes to refine regression, tackling the challenge of distinguishing overlapping wheat grains. Experimental results indicate that Star-YOLO attains a mean Average Precision (mAP) of 97.3% in wheat grain detection, outperforming models like YOLOv5n and RT-DETR in both accuracy and efficiency, with a reduced parameter count. Tests on embedded devices reveal a 36.8% performance enhancement over the baseline model, fulfilling the demands of real-time detection. The Star-YOLO model markedly enhances both the accuracy and efficiency of wheat grain detection while underscoring the potential of deep learning in agricultural image analysis and smart wheat industry advancements.