Abstract
Tea tree seeds are highly sensitive to dehydration and cannot be stored for extended periods, making surface defect detection crucial for preserving their germination rate and overall quality. To address this challenge, we propose Cgc-YOLO, an enhanced YOLO-based model specifically designed to detect small-scale and complex surface defects in tea seeds. A high-resolution imaging system was employed to construct a dataset encompassing five common types of tea tree seeds, capturing diverse defect patterns. Cgc-YOLO incorporates two key improvements: (1) GhostBlock, derived from GhostNetV2, embedded in the Backbone to enhance computational efficiency and long-range feature extraction; and (2) the CPCA attention mechanism, integrated into the Neck, to improve sensitivity to local textures and boundary details, thereby boosting segmentation and localization accuracy. Experimental results demonstrate that Cgc-YOLO achieves 97.6% mAP50 and 94.9% mAP50-95, surpassing YOLO11 by 2.3% and 3.1%, respectively. Furthermore, the model retains a compact size of only 8.5 MB, delivering an excellent balance between accuracy and efficiency. This study presents a robust and lightweight solution for nondestructive detection of tea seed defects, contributing to intelligent seed screening and storage quality assurance.