支持向量机的动物血液光谱特征提取和识别分类

发布时间：2018-06-29 18:51

  本文选题：动物血液 + 荧光光谱　； 参考：《光谱学与光谱分析》2017年12期

【摘要】：利用光谱检测和数据挖掘实现不同种类动物血液光谱数据的精确识别与分类具有重要意义,目前尚未见到较为完善及普适的相关研究报道。实验采集了鸽、鸡、鼠、羊四种动物全血和红细胞溶液(浓度为1%)的荧光光谱数据;基于小波变换的软阈值去噪方法,首先对原始光谱数据进行去噪处理,并确定了717个原始特征(包括荧光峰强度值、荧光峰连线斜率等4类特征);提出以"区分度统计量"为核心的特征提取方法,结合主成分分析法和平均影响值算法,实现了对717个原始特征到2个识别特征的高效筛选;进一步建立了径向基核函数的支持向量机分类器,对四类不同动物的全血荧光光谱数据实现了准确率为100%的识别分类,对红细胞荧光光谱数据实现了94.69%~99.12%的识别率;最后蒙特卡洛交叉验证的结果表明所提出的思路和方法对于动物全血溶液的识别分类具有较好的泛化能力,能对荧光光谱数据进行准确的识别分类,因此能够在进出口检查、食品安全、医药等领域发挥重要作用。针对动物血液荧光光谱,提出的基于"区分度统计量"的特征提取方法,相比于传统的人为特征选取方法,能够从大量原始特征中自动提取少量且有效的识别特征,具有较强的普适性和高效性,为其他领域的光谱特征提取和识别分类提供了一种新的思路。
[Abstract]:It is of great significance to use spectral detection and data mining to accurately identify and classify the blood spectral data of different species of animals. The fluorescence spectrum data of whole blood and red blood cell solution (1%) of pigeon, chicken, mouse and sheep were collected, and the original spectral data were de-noised based on the soft threshold denoising method based on wavelet transform. 717 original features (including fluorescence peak intensity, fluorescence peak line slope, etc.) were determined, and a feature extraction method based on "discriminant statistics" was proposed, which combined principal component analysis (PCA) and average influence value (AIA) algorithm. A highly efficient selection of 717 original features to 2 recognition features is realized, and a support vector machine classifier based on radial basis function (RBF) kernel function is further established. The classification accuracy of 100% is achieved for the whole blood fluorescence spectrum data of four different kinds of animals. The recognition rate of 99.12% was achieved for the red blood cell fluorescence spectrum data. Finally, the results of Monte Carlo cross-validation showed that the proposed method had a good generalization ability for the recognition and classification of animal whole blood solution. Fluorescence spectrum data can be accurately identified and classified, so it can play an important role in import and export inspection, food safety, medicine and other fields. Compared with the traditional artificial feature selection method, the proposed feature extraction method based on discriminant statistics for animal blood fluorescence spectrum can automatically extract a small number of effective recognition features from a large number of original features. It has strong universality and high efficiency, which provides a new way for spectral feature extraction and recognition and classification in other fields.
【作者单位】： 长春理工大学理学院;中国农业科学院长春兽医研究所;西安交通大学数学与统计学院;
【基金】：国家自然科学基金项目(1120420,11426045)资助
【分类号】：O657.3;S852.2
，

本文编号：2083031