光谱学与光谱分析, 2019, 39 (8): 2522, 网络出版: 2019-09-02   

联合光谱-空间信息的短波红外高光谱图像茶叶识别模型

Study on the Tea Identification of Near-Infrared Hyperspectral Image Combining Spectra-Spatial Information
作者单位
1 河南工程学院土木工程学院, 河南 郑州 451191
2 河南工程学院人文社会科学学院, 河南 郑州 451191
3 中国矿业大学(北京)地球科学与测绘工程学院, 北京 100083
4 北京农业信息技术研究中心, 北京 100097
5 国家农业信息化工程技术研究中心, 北京 100097
6 农业部农业信息技术重点实验室, 北京 100097
摘要
茶叶种类识别和等级划分的实践意义重大。 成像光谱技术较传统检测、 识别手段具有图谱合一及快速无损等优势。 获取了君山银针、 无锡白茶、 信阳毛尖、 和六安瓜片4种外观相近的线条形茶叶的短波红外(1 000~2 500 nm)高光谱图像。 首先利用最小噪声分数(MNF)和非参数权重特征提取(NWFE)将高维高光谱数据投影到低维子空间, 然后用单因素方差分析(ANOVA)重新评估投影特征的可分性并选择对茶叶识别较为有效子空间, 同时考虑到“光谱和特征”能较好地表征物质反射属性, 将选择的投影子空间MNF1, MNF2, MNF4, MNF6, MNF8, NWFE1, NWFE2, 及“光谱和特征”一起作为光谱特征集并用SVM分类器获得光谱特征下像元的分类结果。 另一方面, 利用图像本质分解(IID)算法将高光谱图像的光谱分解为自身反射光谱R与阴影成分S; 在均质性较优的光谱范围(1 006~1 900 nm)按照光谱距离对R求取梯度图像并用分水岭算法实现了图像空间分割, 得到空间相关度较高的分割子块。 最后, 将像元分类和图像分割结果进行融合, 具体: 在每个图像分割子块中, 重新统计像元分类结果并按照最大投票法对整个子块的类别进行赋值, 也即联合光谱-空间信息的茶叶识别模型。 结果表明, 构建的模型对4种茶叶的识别结果较为满意, 在仅为约1%水平的训练样本下, 茶叶的总体分类精度达94.3%, Kappa系数为0.92。 该模型还较好地克服了茶叶光谱的“同物异谱”现象, 并期待方法对实践生产具有指导意义。
Abstract
The class-identification and grade-determination of tea have practical importance. Hyperspectral imaging possesses conspicuous advantages in data form of the combination of image and spectra, as well as in fast and undamaged checking in food safety, compared with traditional methods. In this study, hyperspectral images of four kinds of tea which have similar appearance were obtained at the spectral range from 1 000 to 2 500 nm. MNF (Minimum Noise Fraction) and NWFE (Nonparametric Weighted Feature Extraction) were used to rotate and project the hyperspectral data from high dimension to lower subspaces. Then, ANOVA (Analysis of Variance) was used to estimate and select the projected subspaces which have better separability, and they are MNF1, MNF2, MNF4, MNF6, MNF8, NWFE1, NWFE2. Then selected subspaces together with the sum of all original bands were fed to SVM classifier. On the other hand, Finally, IID (image intrinsic decomposition) was applied to decompose the original spectra into material reflectance spectra R and shadow spectra S. Next, gradient image was obtained from R, and watershed algorithm was adapted to segment image in spatial dimension. Finally, results of both pixel-classification and spatial segmentation were fused to have better tea identification. The proposed method was proved to have a satisfying result with an overall accuracy of 94.3% and Kappa coefficient of 0.92 given the only 1% training pixels of all the tea pixels. The proposed model well avoids the phenomenon of same material but different spectra, and significance of reference in practical production is expected.

蔡庆空, 李二俊, 蒋金豹, 乔小军, 蒋瑞波, 冯海宽, 刘绍堂, 崔希民. 联合光谱-空间信息的短波红外高光谱图像茶叶识别模型[J]. 光谱学与光谱分析, 2019, 39(8): 2522. CAI Qing-kong, LI Er-jun, JIANG Jin-bao, QIAO Xiao-jun, JIANG Rui-bo, FENG Hai-kuan, LIU Shao-tang, CUI Xi-min. Study on the Tea Identification of Near-Infrared Hyperspectral Image Combining Spectra-Spatial Information[J]. Spectroscopy and Spectral Analysis, 2019, 39(8): 2522.

本文已被 3 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!