基于深层信念网络的太赫兹光谱识别研究
发布时间：2018-11-24 12:20
【摘要】：太赫兹技术是近年来发展极其迅速的一门新兴交叉学科,由于其拥有很多独特优点,引起国内外研究人员的广泛关注。目前已有一部分太赫兹产品得到了实际运用,展现出极高的使用价值和广阔的应用前景。由于太赫兹光谱具有的“指纹谱”特性,使得对太赫兹光谱的识别成为了鉴别物质、特别是大分子物质(如毒品、中药材、农作物、食品、爆炸物等)的一种重要手段。当前,随着太赫兹时域光谱技术的迅猛发展,仪器设备检测精度以及使用便捷性的提高,物质的太赫兹光谱样本数量呈现出急剧增长的趋势,如何有效的利用这些太赫兹光谱数据对物质进行识别是该领域面临的一个较大的问题。相比于人工定义、提取特征的方法,深层信念网络是从2006年开始重新兴起的一种自动学习特征的方法,而且能更有效的处理规模化数据,其在语音、图像以及自然语言处理领域已得到了成功的应用。针对部分物质在太赫兹波段内没有明显的吸收峰,难以人工定义、提取特征及分类识别的问题。文中结合深层信念网络和KNN分类器的优点,探讨了一种基于深层信念网络的太赫兹光谱识别方法。首先利用S-G滤波和三次样条插值对不同物质的太赫兹透射光谱进行归一化处理；然后由两层受限波尔兹曼机构建深层信念网络模型,并采用逐层无监督的方法来训练模型,以自动提取太赫兹光谱特征；最后用KNN分类器对不同物质的太赫兹透射光谱进行识别。针对目前几乎所有的太赫兹光谱数据库基本只能提供简单的列表查询、名称查询功能,即使用一些关键词来检索对应的光谱,缺少通过光谱检索光谱的功能的问题。本文探索了一种基于局部敏感哈希算法的太赫兹光谱数据库的构建方法,同时,对谱检谱的方法进行了研究,并搭建了一个太赫兹光谱识别的原型系统,对文中探讨方法的有效性进行了验证。
[Abstract]:Terahertz technology is a new interdisciplinary subject which has been developing rapidly in recent years. Because of its many unique advantages, terahertz technology has attracted wide attention of researchers at home and abroad. At present, some terahertz products have been used in practice, showing very high use value and broad application prospect. Because of the "fingerprint spectrum" characteristic of terahertz spectrum, the identification of terahertz spectrum has become an important means of identifying substances, especially macromolecules (such as drugs, Chinese medicinal materials, crops, food, explosives, etc.). At present, with the rapid development of terahertz time-domain spectroscopy technology and the improvement of the precision and ease of use of instruments and equipment, the number of terahertz spectrum samples of materials has shown a trend of rapid growth. How to effectively use these terahertz spectral data to identify matter is a big problem in this field. Compared with the manual definition, the deep belief network is an automatic feature learning method that has been emerging since 2006, and it can deal with the large-scale data more effectively. The field of image and natural language processing has been successfully applied. Because there is no obvious absorption peak of some substances in terahertz band, it is difficult to manually define, extract features and identify classification. Combining the advantages of deep belief network and KNN classifier, a terahertz spectral recognition method based on deep belief network is discussed. First, S-G filter and cubic spline interpolation are used to normalize the terahertz transmission spectra of different materials. Then the deep belief network model is built by two-layer constrained Boltzmann mechanism and the unsupervised method is used to train the model to extract terahertz spectrum feature automatically. Finally, the terahertz transmission spectra of different substances are identified by KNN classifier. At present, almost all terahertz spectral databases can only provide simple list query and name query function, even though some keywords are used to retrieve the corresponding spectrum, it lacks the function of spectrum retrieval through spectrum. In this paper, a method of constructing terahertz spectral database based on local sensitive hashing algorithm is explored. At the same time, the method of spectrum detection is studied, and a prototype system of terahertz spectral recognition is built. The validity of the method discussed in this paper is verified.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：O441.4;O433
本文编号：2353687
[Abstract]:Terahertz technology is a new interdisciplinary subject which has been developing rapidly in recent years. Because of its many unique advantages, terahertz technology has attracted wide attention of researchers at home and abroad. At present, some terahertz products have been used in practice, showing very high use value and broad application prospect. Because of the "fingerprint spectrum" characteristic of terahertz spectrum, the identification of terahertz spectrum has become an important means of identifying substances, especially macromolecules (such as drugs, Chinese medicinal materials, crops, food, explosives, etc.). At present, with the rapid development of terahertz time-domain spectroscopy technology and the improvement of the precision and ease of use of instruments and equipment, the number of terahertz spectrum samples of materials has shown a trend of rapid growth. How to effectively use these terahertz spectral data to identify matter is a big problem in this field. Compared with the manual definition, the deep belief network is an automatic feature learning method that has been emerging since 2006, and it can deal with the large-scale data more effectively. The field of image and natural language processing has been successfully applied. Because there is no obvious absorption peak of some substances in terahertz band, it is difficult to manually define, extract features and identify classification. Combining the advantages of deep belief network and KNN classifier, a terahertz spectral recognition method based on deep belief network is discussed. First, S-G filter and cubic spline interpolation are used to normalize the terahertz transmission spectra of different materials. Then the deep belief network model is built by two-layer constrained Boltzmann mechanism and the unsupervised method is used to train the model to extract terahertz spectrum feature automatically. Finally, the terahertz transmission spectra of different substances are identified by KNN classifier. At present, almost all terahertz spectral databases can only provide simple list query and name query function, even though some keywords are used to retrieve the corresponding spectrum, it lacks the function of spectrum retrieval through spectrum. In this paper, a method of constructing terahertz spectral database based on local sensitive hashing algorithm is explored. At the same time, the method of spectrum detection is studied, and a prototype system of terahertz spectral recognition is built. The validity of the method discussed in this paper is verified.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：O441.4;O433
【参考文献】
相关期刊论文 前2条
1 蒋玉英;葛宏义;廉飞宇;张元;夏善红;;基于THz技术的农产品品质无损检测研究[J];光谱学与光谱分析;2014年08期
2 赵国忠;;太赫兹光谱和成像技术及其应用[J];现代科学仪器;2006年05期
相关硕士学位论文 前1条
1 牧凯军;爆炸物的太赫兹光谱研究[D];首都师范大学;2008年
，本文编号：2353687
本文链接：https://www.wllwen.com/kejilunwen/dianzigongchenglunwen/2353687.html
