Authors

Abstract

Efficient classification of tobacco coated seeds is critical for seed quality assurance in the tobacco industry, yet the nearly identical appearance of coated seeds renders manual sorting labor-intensive and unreliable. Existing automated solutions either lack the precision required for distinguishing sub-millimeter size differences or are too computationally demanding for deployment on edge devices in seed processing facilities. To bridge this gap, this study proposes a lightweight classification model, EMC-YOLO, based on an improved YOLOv10n. This model optimizes YOLOv10n in three aspects: first, replacing the original backbone network with EfficientViT, balancing low computational requirements and global feature extraction capabilities; second, introducing an improved multi-scale spatial pyramid attention (MSPA) into the neck network, utilizing parallel dilated convolution to optimize feature fusion; third, employing the CARAFE operator for upsampling to enhance the edge restoration effect of small targets. Experiments show that the precision, recall, and mAP@0.5 of EMC-YOLO reach 95.12%, 91.20%, and 93.01%, respectively, representing improvements of 2.12, 1.20, and 1.01 percentage points over the original model. Simultaneously, the model reduces the number of parameters by 71%, computational load by 75%, and weight volume by 68.1%. This model achieves high-performance detection with low resource consumption, providing a technical reference for automated seed quality inspection. Validation on a designed seed detection platform under simulated production conditions yielded a coating qualification rate of 95.50%, meeting the 95% industrial benchmark, and confirming the practical feasibility of the proposed method for automated quality inspection in tobacco seed processing lines.

Abstract in Chinese

烟草包衣种子的精准分级对保障种子质量至关重要，然而包衣种子外观高度相似，人工分选劳动密集且易出错。现有自动化方案或精度不足以区分亚毫米级尺寸差异，或计算量过大难以部署于种子加工场景中的资源受限边缘设备。为此，本研究提出一种基于改进YOLOv10n的轻量级分类模型EMC-YOLO。该模型通过三方面优化YOLOv10n：第一，使用EfficientViT替换原主干网络，兼顾低计算量与全局特征提取能力；第二，在颈部网络引入改进的多尺度空间金字塔注意力（MSPA），利用并行空洞卷积优化特征融合；第三，采用CARAFE算子进行上采样，提升小目标的边缘恢复效果。实验显示，EMC-YOLO的精确率、召回率及mAP@0.5分别达到95.12%、91.20%和93.01%，较原模型提升了2.12、1.20和1.01个百分点。同时，模型参数量减少71%，计算量降低75%，权重体积缩小68.1%。该模型能在低资源消耗下实现高性能检测，为种子自动化质量检测提供了技术参考。在设计的种子检测平台上进行模拟生产条件下的验证，包衣合格率达95.50%，超过95%的行业基准，验证了所提方法在烟草种子自动化质量检测中的实际可行性。