基于近红外光谱法的精米品种判别优化研究

Optimization of Polished Rice Varieties Discrimination Based on Near Infrared Spectroscopy

由于地理标志性大米巨大的市场价值,导致掺假行为时有发生。因此,为保证地理标志性大米品牌效益和消费者权益,实现准确的精米品种判别具有重要意义。近红外光谱法是精米品种判别的常用方法,通过提取不同品种近红外光谱中的差异性特征实现品种分类。然而,现有研究中存在特征波长选取性能不足和针对指定品种判别准确度不足的问题,限制了基于近红外光谱法的精米品种判别准确度的提升。针对上述问题,面向东北地区五常、响水、越光、银水四种大米,从特征波长选取和品种判别策略两个方面,研究基于近红外光谱法的精米品种判别优化。首先,为提高特征波长选取性能,将排列熵和自适应滑动窗口分割相结合,提出了基于自适应滑动排列熵的精米光谱特征波长选取方法,并开展与传统特征波长选取算法的对比实验。其次,为提高面向不同指定品种的判别准确度,提出基于判别目标的精米品种判别策略,通过研究光谱预处理算法与分类建模算法匹配优化,建立“指定品种-选定模型-选定算法的”的判别流程。实验结果表明,采用所提出的自适应滑动排列熵算法进行特征波长选取,相比于传统算法精米品种判别误差至少可降低50%;采用所提出的基于判别目标的精米品种判别策略,相比于传统的基于固定模型的判别策略的判别准确度至少可提高2.5%。

Due to the enormous market value of geographically iconic,adulteration or fraud often occurs.Therefore,to ensure the brand benefits from geographically iconic rice and consumer rights,it is important to identify polished rice varieties accurately.Near-infrared spectroscopy is a common method to distinguish polished rice varieties.The varieties can be classified by extracting the different features of different types in near-infrared spectroscopy.However,there are some problems in the existing studies,such as insufficient characteristic wavelength selection performance and insufficient discrimination accuracy for specific varieties,which limit the improvement of the discrimination accuracy of polished rice varieties based on the near-infrared spectroscopy method.In response to the above problems,this paper studies the optimisation of milled rice variety identification based on near-infrared spectroscopy from the two aspects of characteristic wavelength selection and variety identification strategies for four types of rice,Wuchang,Xiangshui,Koshihikari,and Yinshui in Northeast China.First,permutation entropy(PE)and adaptive sliding window(ASW)were combined to improve the feature wavelength selection performance.An adaptive sliding permutation entropy(ASW-PE)based method for selecting the characteristic wavelength of polished rice spectrum was proposed and compared with the traditional algorithm.Secondly,a discriminant strategy based on the discriminant objective was proposed to improve the discriminant accuracy of different specified varieties.By studying the matching optimisation of the spectral preprocessing algorithm and classification modelling algorithm,a discriminant process of“specified cultivation-selected model-selected algorithm”was established.Experimental results show that using the adaptive sliding permutation entropy algorithm proposed in this article to select characteristic wavelengths can reduce the milled rice variety discrimination error by at least 50%compared with the traditional algorithm;using the milled rice variety judge strategy based on the discrimination target proposed in this article.Compared with the conventional judge strategy based on fixed models,the discrimination accuracy can be improved by at least 2.5%.