MCG-YOLOv8n: An Enhanced YOLOv8n Model for Segmentation of Adherent Buckwheat Seeds

To address the challenge of detecting adherent seeds in images of dehuller discharge from buckwheat processing, this study proposes an innovative approach using an enhanced YOLOv8n architecture, termed MCG-YOLOv8n. This model is specifically tailored for precise segmentation of adherent buckwheat seed objects. It incorporates the MobileViT v3 Block into the YOLOv8n-seg backbone network to enhance the feature extraction capability in densely distributed seed areas. At the feature fusion layer, the model incorporates the Generalized Feature Pyramid Network (GFPN) structure for cross-scale feature fusion, coupled with Coordinate Attention (CA) to elevate the detection accuracy of smaller targets. Additionally, the Minimum Point Distance Intersection over Union (MPDIoU) was employed to strengthen the model's generalization ability and mitigate the issue of missed detections caused by dense adherent seeds. Experimental results demonstrate the superior performance of the MCG-YOLOv8n model, achieving high precision (94.70%), recall (93.20%), and mean average precision (mAP@0.5: 95.20%), outperforming other mainstream segmentation models. The analysis of the segmentation results for images, with seed counts ranging from 66 to 435 and a size of 512× 512 pixels, revealed a weighted error rate of 0.70% in counting, with an average detection time of 0.017 s per image. Consequently, the MCG-YOLOv8n model can accurately and rapidly detect high-throughput adherent buckwheat seeds, providing substantial support for the development of online detection systems and deployment on portable mobile devices. © 2013 IEEE.