中文摘要：

提出了一种利用多光谱图像颜色特征进行茶叶分类的新方法,对两种颜色几乎一样用肉眼几乎不能分辨的茶叶进行了分类。图像由MS3100-3CCD光谱成像仪和普通数码相机同时获得,光谱成像仪提供3个波段的图像,由近红外（NIR）、红色（R）和绿色（G）组成,因此它比普通数码照相机包含更丰富的信息,特别NIR波段的图像对有机物的颜色比可见光敏感。提取3CCD光谱成像仪和普通数码照相机各个波段图像颜色的特征即像素偏方差值和平均值进行统计分析,用多光谱图像的NIR图像所提供颜色信息能够辨别这两种颜色几乎一样的茶叶,而普通数码相机无法提供信息进行识别。然后应用人工神经网络技术,对NIR图像像素偏方差值和平均值这两个参数进行建模,建模样本40个,每个样本为20个,预测样本20个,每个样本为10个。结果表明,在阈值为0.3,对两种茶叶进行分类得到了100%识别率,此研究为茶叶的分类提供一种快速和无损的新方法。

英文摘要：

Tea is one of the most popular beverages worldwide.Its categories have a great relationship to its beneficial medicinal properties.The present work attempted to study the feasibility to use multispectral imaging technique as a rapid and non-destructive method to discriminate tea varieties.Two categories of tea discriminated hardly by naked eye were sorted.The images were 1 036 pixels vertically by 1 384 pixels horizontally with 24-bit depth,and were captured using a red（R） waveband,near infrared（NIR） waveband and green（G） waveband multispectral digital imager,MS3100（Duncan Technologies,Inc.,CA,USA）.The three wavebands of image（Red,Green,NIR） can be composed into one image which contains more information than images recorded by ordinary digital cameras,especially,the NIR image is more sensitive to the color of organic matter than visible spectrum.The three images of one sample can be obtained simultaneously.The color features of tea were calculated using the standard notations： mean and mean square deviation.Then,the two color features of 3CCD and ordinary digital cameras were extracted and calculated by Matlab 7.3 software respectively,and were contrasted.A total of 60 samples were adopted,and the features of mean and mean square deviation of NIR waveband image were applied as inputs to a back propagation neural network（BP-ANN） with one hidden layer.The forty samples（twenty for each category） were selected randomly to build BP-ANN model,and this model was used to predict the varieties of 20 unknown samples（ten for each category）.The two categories of tea can be discriminated by the information of color of images of 3CCD,but can not by the ordinary digital cameras.The result indicted that the discrimination rate of classification set of BP-ANN model was up to 100% within 0.3 of threshold.It concluded that multi-spectral imaging technique has a high potential to identify categories of green tea fast and non-destructively.