Authors

Abstract

The quality inspection of sorghum grains after drying is critical for ensuring storage stability and processing quality. However, traditional manual inspection methods are inefficient, time-consuming, and prone to subjective biases. To address these challenges, this study proposes an improved YOLOv8s model, named YOLOv8s-GBD, for accurate sorghum grain classification and detection. The proposed model integrates three key enhancements. First, the DFC Attention-based bottleneck structure from GhostNetV2 is incorporated into the C2f module to enhance global context awareness and improve detection accuracy. Second, a bi-directional feature pyramid network (BiFPN) is introduced to optimize multi-scale feature fusion. Third, a dynamic head framework based on attention mechanisms is adopted to strengthen feature representation, further boosting the model’s accuracy and robustness. Experimental results demonstrate that the YOLOv8-GBD model achieves 95.2% precision, 98.0% recall, 98.7% mAP1, 97.5% mAP2, and 96.6% F1-score. Compared to the original YOLOv8s model, these metrics show improvements of 3.1%, 1.7%, 1.2%, 1.0%, and 2.4%, respectively. Furthermore, the YOLOv8-GBD model outperforms other YOLO series models in sorghum grain classification and detection tasks. In conclusion, the YOLOv8s-GBD model meets the requirements for high accuracy in sorghum grain quality inspection, offering a robust solution for practical applications.

Abstract in Chinese

干燥后高粱籽粒的品质检测对保障仓储稳定性与加工品质至关重要。然而传统人工检测方式效率低下、耗时长，且容易存在主观偏差。针对上述问题，本文提出一种改进 YOLOv8s 模型 YOLOv8s-GBD，实现高粱籽粒精准分类与检测。该模型包含三项核心改进：第一，在 C2f 模块中嵌入 GhostNetV2 的 DFC 注意力瓶颈结构，增强模型全局上下文感知能力，提升检测精度；第二，引入双向特征金字塔网络（BiFPN）优化多尺度特征融合；第三，采用基于注意力机制的动态检测头框架强化特征表征能力，进一步提升模型精度与鲁棒性。实验结果表明：YOLOv8s-GBD 模型精确率 95.2%、召回率 98.0%、mAP1 为 98.7%、mAP2 为 97.5%、F1 分数 96.6%。相较于原始 YOLOv8s 模型，各项指标分别提升 3.1%、1.7%、1.2%、1.0%、2.4%。同时，在高粱籽粒分类检测任务中，YOLOv8s-GBD 模型性能优于其他 YOLO 系列模型。综上所述，YOLOv8s-GBD 模型能够满足高粱籽粒品质检测的高精度需求，可为实际工程应用提供可靠方案。