高光谱技术的羊肉品种多波段识别研究

Study on Multi-Bands Recognition for Varieties of Mutton by Using Hyperspectral Technologies

利用可见/近红外（400-1000 nm）及近红外（900-1700 nm）高光谱成像技术对宁夏地区滩寒杂交、盐池滩羊、小尾寒羊三个品种羊肉进行识别研究。针对不同波段光谱特点,分别优选出 Baseline 及 SG卷积平滑光谱预处理方法,并运用连续投影算法（SPA）提取特征波长,结合线性判别分析（LDA）及径向基核函数支持向量机（RBFSVM）模型进行全波段及特征波长识别分析。结果表明不同波段高光谱对羊肉品种识别均获得较好效果,其中400-1000 nm波段采用Baseline-Fullwave-RBFSVM及12个特征波长下准确率为100%与98.75%,900-1700 nm 波段采用 Baseline-Fullwave-RBFSVM 及7个特征波长下准确率为96.25%与87.80%；RBFSVM非线性分类准确率高于 LDA线性判别结果,400-1000 nm波段识别准确率优于900-1700 nm波段,说明三种羊肉在色泽纹理上差异比成分含量显著,利用高光谱成像技术结合RBFSVM方法能够获得较优的羊肉品种识别效果。

This paper focused on the research on identifying and classifying for mutton varieties of Tan-han hybrid sheep,Yanchi Tan-sheep and small-tailed sheep in Ningxia by using visible/near-infrared （400-1 000 nm）.Near infrared （900-1 700 nm） hyperspectral technologies,baseline and SG convolution smoothing spectra pretreatment methods were applied respectively ac-cording to the characteristics of different spectrum bands;the characteristic wavelengths were extracted by using successive pro-j ection algorithm （SPA）;then combined with linear discriminant analysis （LDA）and radial basis kernel function of support vec-tor machine （RBFSVM）model were applied to identify the different mutton varieties under characteristic wavelengths and full-wave bands.Results showed that there were good effects for mutton varieties identification in different hyperspectral bands, among which Baseline-Fullwave-RBFSVM and the same models under 12 characteristic wavelengths obtained accuracy of 100%and 98.75% in 400-1 000 nm respectively,and Baseline-Fullwave-RBFSVM and the same models under 7 characteristic wave-lengths obtained accuracy of 96.25% and 87.80% in 900-1 700 nm respectively.The identification accuracy of RBFSVM non-linear classification was higher than the LDA linear discriminant result,meanwhile the identification accuracy in 400-1 000 nm bands was better than in 1 000-1 700 nm bands,which explained that the differences of color and texture were more significant than the component contents among the 3 varieties mutton.Combined hyperspectral technologies with RBFSVM models can ob-tain a better recognition effect of mutton varieties.