小麦籽粒外观品质检测装置设计与试验

Design and Experiment of Detection Device for Wheat Kernel Appearance Quality

摘要:

针对小麦籽粒外观品质传统人工检测方法效率低、标准不统一,现有检测设备功能单一、操作复杂且成本高昂的问题,本研究提出了小麦不完善粒自动识别及籽粒多表型参数精准获取方法。设计了小麦籽粒外观品质检测装置结构、工作流程及控制方案,构建了包含6类小麦样本（正常粒和5类不完善粒）的数据集。通过更换骨干网络及卷积模块,提出了基于轻量化深度学习网络的小麦不完善粒识别方法,在多数分类数据集上精确率及召回率均达到95%以上。对籽粒图像进行畸变校正处理,消除畸变对长宽比测量的影响,采用边缘检测及分水岭算法对籽粒轮廓进行分割,构建了基于图像处理的小麦籽粒多表型参数获取模型,粒长、粒宽及长宽比均方根误差分别为0.19 mm、0.14 mm和0.10。最后,集成了系统样机,对小麦籽粒外观品质检测装置的试验结果表明,小麦不完善粒识别及多表型参数获取模型具有较高的识别及测量精度,样机具有运行稳定、成本效益良好、检测参数全面的综合优势,研究结果可为小麦优良品种快速选育及性状改良提供参考。

Abstract:

Aiming to address the problems of low efficiency and inconsistent standards in traditional manual inspection of wheat grain appearance quality,as well as the limitations of existing detection equipment such as single functionality,complex operation,and high cost,an automatic identification method for imperfect wheat grains and an accurate acquisition method for multiple phenotypic parameters of grains were proposed. The structure,workflow,and control scheme of a wheat grain appearance quality detection device were designed. A dataset containing six types of wheat samples (normal grains and five types of imperfect grains) was constructed. By replacing the backbone network and convolution modules,an identification method for imperfect wheat grains based on a lightweight deep learning network was proposed,with both precision and recall exceeding 95% on most classification datasets. Distortion correction was applied to grain images to eliminate the influence of distortion on aspect ratio measurement. Edge detection and watershed algorithms were used to segment grain contours,and an image-processing-based model for acquiring multiple phenotypic parameters of wheat grains was established. The root mean square errors of grain length,grain width,and aspect ratio were 0.19 mm,0.14 mm,and 0.10,respectively. Finally,a prototype system was integrated. Experimental results of the wheat grain appearance quality detection device showed that the proposed imperfect wheat grain identification model and the multiple phenotypic parameter acquisition model achieved high recognition and measurement accuracy. The developed prototype demonstrated stable operation,good cost-effectiveness,and comprehensive detection parameters. The results can provide a reference for rapid breeding of high-quality wheat varieties and trait improvement.