基于LBP和深度学习的人脸特征提取
发布时间：2019-02-15 12:28
【摘要】：在当今信息化时代,如何准确鉴定一个人的身份、保护信息安全,已成为一个必须解决的关键社会问题。人脸识别技术是最有发展潜力的生物特征识别技术之一,由于其具有简单直观、不易复制、安全性高、操作容易等特点而在身份验证和识别场合具有巨大的应用价值和广阔的应用前景。而特征提取的好坏对人脸的正确识别有着至关重要的影响,因此,如何提取稳定有效的人脸特征,使得提取的特征尽可能多的含有有利于识别的类别信息,以及如何将多种不同的特征相结合来实现更为理想的分类结果等都是现阶段人脸识别的研究热点。通过对现有人脸识别相关文献的阅读,本文在总结前人研究成果的基础上,深入研究了基于局部二值模式和深度学习的特征提取方法,并在此基础上做了以下工作:(1)针对选用一种特征不足以捕捉人脸图像多方面的识别信息问题,综合考虑局部二值模式(LBP)的改进算法ELBP与离散余弦变换(DCT)的优缺点,提出了一种将ELBP与离散余弦变换(DCT)相结合来进行特征提取的方法,该方法将人脸图像经DCT变换后的少量低频系数作为人脸的频域特征,将人脸图像中眼部和嘴部区域的ELBP特征作为人脸的空域特征,并使用PCA方法对所提取的空频域特征进行有效融合,得到更有效的人脸特征,通过在ORL人脸库和Yale人脸库上的实验验证了该方法的有效性。(2)针对深度信念网络(DBNs)忽略了图像局部结构,难以学习到人脸图像的局部特征以及网络训练时间过长等问题进行了研究,提出将LBP特征作为DBNs的输入,并在DBNs的训练过程中引入极限学习机(ELM)来加快DBNs的训练速度,最后用训练好的网络进行分类识别。在ORL人脸库和FERET人脸库上对不同样本规模和不同分辨率的图像进行实验,结果表明:与单独采用LBP或DBNs提取特征的方法相比,该方法取得了较好的学习效率和识别效果。
[Abstract]:In today's information age, how to accurately identify a person and protect information security has become a key social problem that must be solved. Face recognition is one of the most promising biometric recognition techniques. Because of its easy operation, it has great application value and broad application prospect in authentication and identification. The quality of feature extraction plays an important role in correct face recognition. Therefore, how to extract stable and effective face features makes the extracted features contain as many kinds of information as possible. And how to combine a variety of different features to achieve a more ideal classification results are now the focus of face recognition research. On the basis of summarizing the previous research results, this paper deeply studies the feature extraction method based on local binary pattern and depth learning. On this basis, the following work has been done: (1) aiming at the problem of selecting a feature that is not sufficient to capture face image in many aspects, Considering the advantages and disadvantages of the improved algorithm of local binary mode (LBP) (ELBP) and discrete cosine transform (DCT), a method of feature extraction by combining ELBP with discrete cosine transform (DCT) is proposed. In this method, a small amount of low frequency coefficients after DCT transform are used as the frequency domain features of the face, and the ELBP features of the eye and mouth regions in the face image are taken as the spatial features of the face. PCA method is used to effectively fuse the extracted space-frequency domain features to obtain more effective facial features. Experiments on ORL face database and Yale face database demonstrate the effectiveness of the method. (2) the local image structure is ignored in depth belief network (DBNs). It is difficult to learn the local features of face images and the network training time is too long to study. It is proposed that the LBP feature be taken as the input of DBNs, and the extreme learning machine (ELM) is introduced in the process of DBNs training to speed up the training speed of DBNs. Finally, the trained network is used to classify and identify. Experiments on images with different sample sizes and different resolutions are carried out on ORL and FERET face databases. The results show that compared with the methods using LBP or DBNs alone, this method has better learning efficiency and recognition effect.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
[Abstract]:In today's information age, how to accurately identify a person and protect information security has become a key social problem that must be solved. Face recognition is one of the most promising biometric recognition techniques. Because of its easy operation, it has great application value and broad application prospect in authentication and identification. The quality of feature extraction plays an important role in correct face recognition. Therefore, how to extract stable and effective face features makes the extracted features contain as many kinds of information as possible. And how to combine a variety of different features to achieve a more ideal classification results are now the focus of face recognition research. On the basis of summarizing the previous research results, this paper deeply studies the feature extraction method based on local binary pattern and depth learning. On this basis, the following work has been done: (1) aiming at the problem of selecting a feature that is not sufficient to capture face image in many aspects, Considering the advantages and disadvantages of the improved algorithm of local binary mode (LBP) (ELBP) and discrete cosine transform (DCT), a method of feature extraction by combining ELBP with discrete cosine transform (DCT) is proposed. In this method, a small amount of low frequency coefficients after DCT transform are used as the frequency domain features of the face, and the ELBP features of the eye and mouth regions in the face image are taken as the spatial features of the face. PCA method is used to effectively fuse the extracted space-frequency domain features to obtain more effective facial features. Experiments on ORL face database and Yale face database demonstrate the effectiveness of the method. (2) the local image structure is ignored in depth belief network (DBNs). It is difficult to learn the local features of face images and the network training time is too long to study. It is proposed that the LBP feature be taken as the input of DBNs, and the extreme learning machine (ELM) is introduced in the process of DBNs training to speed up the training speed of DBNs. Finally, the trained network is used to classify and identify. Experiments on images with different sample sizes and different resolutions are carried out on ORL and FERET face databases. The results show that compared with the methods using LBP or DBNs alone, this method has better learning efficiency and recognition effect.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 江宇闻;朱思铭;;基于Overcomplete ICA的人脸特征提取[J];计算机科学;2005年07期
2 陈鹏;;人脸特征提取方法的研究[J];中国水运(理论版);2006年12期
3 魏冬冬;谌海新;聂铁铸;;人脸特征提取与识别技术研究[J];计算机与现代化;2007年03期
4 王宏勇;王青青;;三维人脸特征提取方法综述[J];电子科技;2012年12期
5 王fE;杨光;;改进的脉冲耦合神经网络人脸特征提取方法[J];计算机工程与应用;2013年01期
6 宋宇,周激流,何坤,刘智明,黎奎;多尺度人脸特征提取[J];计算机应用研究;2005年04期
7 陶宇权;程德福;李成荣;徐U，
本文编号：2423348
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2423348.html
