基于YOLOv11的杂粮成分识别

Based on YOLOv11 for Multi-Grain Component Recognition

摘要: 杂粮由于其粮食种类丰富、营养全面等优点存在着极大的市场空间。为便于消费者了解杂粮的品质、工商部门检测杂粮的质量以及农民设备的局限性，呼应市场的需求，本文设计了一种基于YOLO模型的谷物分类计数器，且仅用手机相机即可实现数据收集。针对杂粮识别场景中颗粒目标细小、密集易粘连、类别外观高度相似、复杂背景干扰强等问题，本文通过针对性数据增强、优化分类与定位损失权重、引入测试时增强TTA及差异化后处理策略。实验结果表明，改进模型在10类杂粮识别任务中，有8类杂粮都具备较高分类精度与计数稳定性，模型对大多数谷物的识别准确率达到了85%以上。改进模型的平均绝对误差由1.130降至0.880，平均相对误差由3.5%降至2.9%。本文也通过混淆矩阵、性能曲线、对比实验、重复性检验与误差分析验证了鲁棒性与实用性，可有效实现杂粮成分快速判别与精准计数。

Abstract: Coarse cereals hold significant market potential due to their rich variety and comprehensive nutritional value. To help consumers assess product quality, support market regulation, and address the limitations of equipment in agricultural settings, this paper presents a grain classification and counting system based on a YOLO model, which only requires a mobile phone camera for data collection. To address challenges in coarse cereal recognition, such as small and densely packed particles, high inter-class visual similarity, and complex background interference, this study employs targeted data augmentation, optimizes the weights of classification and localization losses, introduces Test Time Augmentation (TTA), and applies differentiated post-processing strategies. Experimental results show that the improved model achieves high classification accuracy and counting stability for 8 out of 10 coarse cereal types, with recognition accuracy exceeding 85% for most grains. The mean absolute error of the enhanced model is reduced from 1.130 to 0.880, and the mean relative error decreases from 3.5% to 2.9%. The robustness and practicality of the system are further validated through confusion matrices, performance curves, comparative experiments, repeatability tests, and error analysis, demonstrating its effectiveness for rapid composition identification and accurate counting of mixed grains.