Authors

Abstract

To address the low inspection efficiency and limited recognition accuracy in sorghum grain quality assessment for brewing enterprises, this study proposes YOLOv11-SPA, an efficient and real-time detection model based on an improved YOLOv11n architecture. First, the space-to-depth convolution module (SPDConv) is introduced into the backbone network to replace conventional convolution blocks, effectively mitigating the loss of spatial information for small targets caused by downsampling operations. Second, the parallelized patch-aware attention (PPA) module is integrated into the neck network to enhance local feature representation and improve the detection of subtle defect features such as moldy and cracked grains. Third, an adaptive threshold focal loss (ATFL) is proposed to dynamically adjust sample weights, improving the model’s discrimination capability for visually similar categories (e.g., grains with husk residue and intact grains). Experimental results on a self-constructed sorghum seed dataset show that YOLOv11-SPA achieves 80.1% Precision, 79.7% Recall, and 85.9% mAP50, outperforming the baseline YOLOv11n by 5.6, 5.9, and 6.2 percentage points, respectively. With only 3.4 M parameters, the proposed model achieves an inference speed of 205 FPS, meeting real-time detection requirements while maintaining high accuracy. These results demonstrate that YOLOv11-SPA provides an effective solution for automated sorghum grain defect inspection and offers promising potential for intelligent quality control in the modern brewing industry.

Abstract in Chinese

针对现阶段酿酒企业高粱籽粒检测效率及识别率较低等问题，提出了一种基于改进YOLOv11n的高效、精准、实时检测模型YOLOv11-SPA。首先，在骨干网络中引入空间到深度卷积模块（SPDConv）替代原始卷积模块，有效缓解因池化操作导致的小目标空间信息丢失问题；其次，在颈部网络嵌入并行化块感知注意力模块（PPA），通过强化局部特征表征能力，提升模型对霉斑籽粒、裂纹籽粒等细微特征的感知能力；最后，设计自适应阈值焦点损失（ATFL）函数，通过动态调整样本权重，优化模型对"包壳残留籽粒"与"完好籽粒"等高相似度样本的区分能力。试验结果表明，该模型在自建高粱籽粒数据集上实现了80.1%精确率、79.7%召回率和85.9%的模型平均精度，较基准模型YOLOv11n分别提升了 5.6、5.9 和 6.2个百分点；最终模型在3.4M参数下实现每秒205帧的推理速度，证明了该模型在保持高精度的同时也满足实时检测需求，对现代化酿造产业具有重要意义。