基于张量分析的玉米种子高光谱图像最优波段选择

发布时间：2018-06-25 10:56

本文选题：玉米种子 + 高光谱图像　；参考：《江南大学》2017年硕士论文

【摘要】：玉米是世界总产量最高的粮食作物,被广泛应用到食品的生产、工业原料的制造以及畜牧业饲料的加工。玉米种子品类的鉴别工作在减少种子混杂现象、保证农业生产的顺利进行方面具有重要价值。高光谱图像技术具有图谱合一的特点,可同时获得玉米种子的图像信息和光谱信息,在玉米种子品种识别中得到了越来越多的重视,并且取得了很高的识别精度。种子分类特征的充分挖掘获取是识别模型精度和鲁棒性的保证。尽管高光谱图像技术可获得种子的图像特征和光谱特征等,但是现有的种子品种识别多是利用高光谱图像的单一光谱特征,导致高光谱图像技术的优势没有被充分利用。另一方面,高光谱图像波段数目众多,给推广应用到种子品种识别在线检测设备上时带来了困难。本文旨在将高光谱图像技术与基于张量(Tensor)分析的多特征波段选择方法相结合,研究一种具有快速性、高准确和高鲁棒性等特点的玉米种子无损检测方法。主要的研究内容包括:1.利用联合偏度算法(JS)选择高光谱图像的最优波段,用于开发种子品类的分级系统。本课题利用高光谱图像采集系统获取17类共1632粒玉米种子在400-1000nm波段范围内的高光谱图像。利用联合偏度算法选择了高光谱图像的最优波段,建立联合特征条件下的最小二乘支持向量机(LS-SVM)种子分类模型。实验结果表明:基于联合偏度的波段选择算法的分类精度要高于无信息变量消除法和连续投影算法,为种子高光谱图像识别技术的准确和快速的识别提供了可行的途径。2.利用有监督的多线性判别分析(MLDA)波段选择算法研究了玉米种子的高光谱图像联合特征的种子分类识别。将MLDA与JS选择后的特征集构建LS-SVM种子分类模型,比较相同条件下的识别精度。实验结果表明:MLDA波段选择方法在相同波段下比JS波段选择方法具有更高的效率。在相同的精度条件下,MLDA波段选择方法可以获得更少的波段数目,这对于开发更高效的种子高光谱图像识别系统是有利的。3.利用多模型与MLDA波段选择算法结合的策略研究了亲缘关系玉米种子的分类鉴选。首先,对874-1734nm波段范围内的两大类亲缘关系玉米品种建立类间切换模型进行初分,再通过构建两个类内的子模型实现细分。同时,为了提高检测的速度和减少模型构建的空间计算量,采用MLDA波段选择方法来选择最优波段。结果表明:多模型在全波段和最优波段下都取得了较高的识别效果,在不同场景下也具有较高的鲁棒性。表明利用多模型和MLDA波段选择方法结合的策略可实现亲缘关系玉米种子高光谱图像的纯度鉴选。
[Abstract]:Corn is the highest grain crop in the world. It is widely used in food production, industrial raw material manufacture and animal husbandry feed processing. The identification of maize seed species has important value in reducing seed mixing and ensuring the smooth progress of agricultural production. The hyperspectral image technology is characterized by the combination of maps and spectra, which can obtain the image information and spectral information of maize seeds at the same time. More and more attention has been paid to the recognition of maize seed varieties, and high recognition accuracy has been obtained. The sufficient mining of seed classification features ensures the accuracy and robustness of the recognition model. Although the hyperspectral image technology can obtain the image features and spectral features of seeds, most of the existing seed varieties recognition is based on the single spectral characteristics of hyperspectral images, resulting in the advantage of hyperspectral image technology has not been fully utilized. On the other hand, the number of bands in hyperspectral images is very large, which makes it difficult to popularize and apply to the on-line detection equipment for seed variety recognition. The aim of this paper is to combine hyperspectral image technology with multi-feature band selection method based on Zhang Liang (Tensor) analysis to study a method of maize seed nondestructive detection with the characteristics of rapidity, high accuracy and high robustness. The main research contents include: 1. The joint bias algorithm (JS) is used to select the optimal bands of hyperspectral images to develop a seed classification system. In this paper, the hyperspectral images of 17 kinds of 1632 corn seeds in 400-1000nm band were obtained by using a hyperspectral image acquisition system. The optimal band of hyperspectral images is selected by using the joint bias algorithm, and the seed classification model of least squares support vector machine (LS-SVM) under the condition of joint feature is established. The experimental results show that the classification accuracy of band selection algorithm based on joint bias is higher than that of non-information variable elimination method and continuous projection algorithm, which provides a feasible way for accurate and fast recognition of seed hyperspectral images. A supervised multilinear discriminant analysis (MLDA) band selection algorithm was used to study the seed classification and recognition of the hyperspectral images of maize seeds. The LS-SVM seed classification model is constructed by using MLDA and JS selected feature sets, and the recognition accuracy is compared under the same conditions. The experimental results show that the proportion of MLDA band selection method is more efficient than JS band selection method in the same band. Under the same precision condition, less band number can be obtained by using MLDA band selection method, which is beneficial to the development of a more efficient seed hyperspectral image recognition system. The classification and selection of related maize seeds were studied by combining multiple models with MLDA band selection algorithm. Firstly, the inter-class switching model was established for two kinds of kinship maize varieties in 874-1734nm band, and then submodels were constructed to realize subdivision. At the same time, in order to improve the speed of detection and reduce the spatial computation of model construction, MLDA band selection method is used to select the optimal band. The results show that the multi-model has a higher recognition effect in both the full and optimal bands, and has a higher robustness in different scenarios. The results showed that the combination of multiple models and MLDA band selection method could be used to identify the purity of the hyperspectral images of related maize seeds.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S513;TP391.41

【参考文献】