Real-Time Detection of Varieties and Defects in Moving Corn Seeds Based on YOLO-SBWL

Sorting corn seeds before sowing is crucial to ensure the varietal purity of the seeds and the yield of the crop. However, most of the existing methods for sorting corn seeds cannot detect both varieties and defects simultaneously. Detecting seeds in motion is more difficult than at rest, and many models pursue high accuracy at the expense of model inference time. To address these issues, this study proposed a real-time detection model, YOLO-SBWL, that simultaneously identifies corn seed varieties and surface defects by using images taken at different conveyor speeds. False detection of damaged seeds was addressed by inserting a simple and parameter-free attention mechanism (SimAM) into the original “you only look once” (YOLO)v7 network. At the neck of the network, the path-aggregation feature pyramid network was replaced with the weighted bi-directional feature pyramid network (BiFPN) to increase the accuracy of classifying undamaged corn seeds. The Wise-IoU loss function supplanted the CIoU loss function to mitigate the adverse impacts caused by low-quality samples. Finally, the improved model was pruned using layer-adaptive magnitude-based pruning (LAMP) to effectively compress the model. The YOLO-SBWL model demonstrated a mean average precision of 97.21%, which was 2.59% higher than the original network. The GFLOPs were reduced by 67.16%, and the model size decreased by 67.21%. The average accuracy of the model for corn seeds during the conveyor belt movement remained above 96.17%, and the inference times were within 11 ms. This study provided technical support for the swift and precise identification of corn seeds during transport.