基于统计机器学习的光谱识别技术

发布时间：2018-08-08 20:23

【摘要】：随着人造卫星、载人航天飞船和国际空间站等航天器的发射,空间目标识别是进一步安全开发空间资源的前提条件。而有效识别空间目标的表面材质对进一步识别目标具有重要的现实意义和应用价值。散射光谱能够有效表征被测样本的表面特性,统计机器学习为解决样本之间不易分类识别的问题,提供了技术手段。本论文基于散射光谱,结合四种统计机器学习算法,研究了空间目标材质的识别问题。具体研究内容如下:1.搭建材质散射光谱测量系统,该系统可测得材质多角度下的的散射光谱。对测得的散射光谱进行了预处理,分别为:去噪、计算双向反射分布函数和归一化三个部分,并建立了材质数据库。2.基于散射光谱与统计机器学习算法原理,分别建立了朴素贝叶斯分类器、K近邻算法、误差反向传播神经网络和卷积神经网络的算法框架,并利用MATLAB软件进行了编程实现。3.基于散射光谱,结合朴素贝叶斯分类器、K近邻算法、误差反向传播神经网络和卷积神经网络四种算法,对经过预处理的材质散射光谱进行了分类识别实验,并对识别结果进行了对比分析。研究结果表明:(1)夹角余弦和欧氏距离结合嵌入使用于K近邻算法时,由于综合考虑了光谱的线型特征与幅度特征,具有精度高,耗时较少的特点,该方法在基于散射光谱的识别领域,具有一定的适用性;(2)卷积神经网络由于特殊的网络结构,具有耗时少,精度较高的特点,该方法对于大数据量的空间目标分类识别领域,具有区别于其他方法一定的优越性和适用性。
[Abstract]:With the launch of satellites, manned spaceships and international space stations, space target recognition is a prerequisite for the further development of space resources. It is very important to recognize the surface material of the space target effectively for the further recognition of the target. The scattering spectrum can effectively characterize the surface characteristics of the measured samples. Statistical machine learning provides a technical means to solve the problem of difficult classification and identification between samples. In this paper, based on scattering spectrum and four statistical machine learning algorithms, the recognition of spatial target materials is studied. The specific contents of the study are as follows: 1. The material scattering spectrum measurement system is set up, which can measure the scattering spectrum of material with multiple angles. The measured scattering spectra are preprocessed into three parts: denoising, calculating the bidirectional reflectance distribution function and normalization, and the material database. 2. Based on the theory of scattering spectrum and statistical machine learning algorithm, the algorithm framework of naive Bayesian classifier K nearest neighbor algorithm, error back propagation neural network and convolution neural network is established, and the program of. 3 is implemented by using MATLAB software. Based on scattering spectrum, combined with naive Bayesian classifier and K-nearest neighbor algorithm, error back-propagation neural network and convolution neural network, the pre-processed material scattering spectrum was classified and recognized. The recognition results are compared and analyzed. The results show that: (1) when the inclusion cosine and Euclidean distance are embedded in the K-nearest neighbor algorithm, because the linear and amplitude characteristics of the spectrum are taken into account, it has the characteristics of high precision and less time consuming. This method has some applicability in the field of recognition based on scattering spectrum. (2) because of the special network structure, the convolution neural network has the characteristics of less time consuming and higher precision. It has some advantages and applicability which is different from other methods.
【学位授予单位】：长春理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：V419;O657.3

【参考文献】