精米品种多特征模型融合分类与外观品质多参数检测应用研究

Study on Multi-Feature Model Fusion Variety Classification and Multi-Parameter Appearance Inspection for Milled Rice

大米是我国最重要的谷类作物。如何准确的实现地理标志性大米的品种鉴别和外观品质评价,不仅关乎消费者切身利益,而且关乎零售商和生产商信誉,是一项广泛关注的问题。首先,为实现集成化的精米品种识别和外观品质检测应用,建立一种精米品种分类与外观品质多参数检测系统。该系统采用近红外光谱仪搭配漫反射附件采集米粉光谱信息,可以实现基于近红外光谱法的精米品种分类;采用互补金属氧化物半导体(CMOS)显微相机配合机电控制系统采集米粒图像,基于图像法实现精米外观品质多参数检测,检测对象包括裂纹、长宽、垩白、碎粒和黄粒。以上述系统为基础,为提高精米品种分类精度,提出一种基于光谱-图像特征模型融合的精米品种分类方法。以近红外光谱特征与多图像特征作为输入参数,以精米品种编号作为输出参数,基于偏最小二乘方法(PLS)建立品种分类融合模型。在不同融合方案的建模过程中,每种融合方案都采用变量投影重要性分析方法(VIP)实现融合模型输入参数的最优筛选,然后通过不同融合方案分类精度对比确定最优融合模型。最后,开展精米外观品质多参数检测和不同精米品种分类方法性能对比实验。结果表明,建立的精米品种分类与外观品质多参数检测系统可以实现包括裂纹米率、粒型、垩白米率、碎米率和黄粒米率的精米外观品质多参数检测,检测精度范围为89.2%~97.0%;提出的基于光谱-图像特征模型融合的精米品种分类方法相比于传统方法可以提高精米品种分类精度,相比于传统方法中效果较好的近红外光谱法,面向五常、响水、银水、越光四种大米的分类精度可提高2.5%~7.5%。

Rice is the most important cereal crop in China.To accurately realize the variety identification and appearance quality evaluation of geographically iconic rice is not only related to consumers'interests but also to the reputation of retailers and manufacturers,which is a widespread concern.Firstly,to realize the integrated application of milled rice variety recognition and appearance quality detection,a multi-parameter detection system for milled rice variety and appearance quality was established.The system uses an NIR spectrometer with a diffuse reflectance accessory to collect the spectral information of rice flour,which can realize the classification of milled rice varieties based on NIR spectroscopy.Multi-parameter detection of milled rice appearance quality was realized based on the image method using the Complementary metal-oxide-semiconductor(CMOS)camera.The detection objects included cracks,length/width,chalkiness,broken grains and yellow grains.Based on the above system,this paper proposed a milled rice variety classification method based on spectral-image feature model fusion to improve the classification accuracy of milled rice varieties.In this method,the NIR spectral features and multi-image features were used as the input parameters,the milled rice variety number was used as the output parameters,and a variety classification fusion model was established based on the Partial least squares(PLS)method.In the modeling process of each fusion scheme,the variable projection importance analysis(VIP)method was used to achieve the optimal selection of the input parameters.Then the optimal fusion model was determined by comparing the classification accuracy of different fusion schemes.Finally,the multi-parameter detection experiment of milled rice appearance quality and the performance comparison experiment of different milled rice variety classification methods were carried out.Experimental results showed that the detection system established in this paper could realize the multi-parameter detection of milled rice appearance quality,including broken rice rate,length-width ratio,fissured rice rate,chalky rice rate,and yellow-colored rice rate,for which the detection accuracy range was 89.2%~97.0%.The proposed milled rice variety classification method based on the spectral-image feature model fusion could improve the classification accuracy of milled rice varieties.Compared with the NIRS method,which has a better effect than the traditional methods,the classification accuracy of Wuchang,Xiangshui,Yinshui,and Yuiguang rice varieties can be improved by 2.5%~7.5%using the new variety classification method.