复杂光照条件下的人脸识别研究
|
摘要
近年来,随着人工智能和生物技术的迅速发展,自动人脸识别已成为图像处理和模式识别中非常活跃的研究课题。
本文在研究了人脸识别的基本理论和关键技术的基础上,重点讨论了不同光照条件下,彩色静止图像的人脸检测和识别问题。主要工作归纳如下:
第一,人脸检测。首先,本文对人脸肤色的聚类特性进行了研究,通过不同颜色空间中肤色聚类效果的比较,进一步改进了HSI颜色模型,构建了H-SI-I颜色模型。其次,自建了具有100张图片的人脸图像库TXK,并通过大量实验建立了H-SI-I肤色模型。最后,对人脸图像库TXK中的图像进行粗检、滤波、细检,最终定位出人脸。实验证明基于H-SI-I肤色模型的人脸检测算法是非常有效的。而且它对光照亮度也具有很强的鲁棒性。
第二,人脸图像预处理。为了提高人脸识别的精度,在人脸识别之前都需要进行人脸图像的预处理工作。预处理包括很多内容,例如对人脸图像进行滤波去噪、灰度转换、灰度归一化、几何归一化等。好的预处理将会直接提高最终的人脸识别率。因此在人脸识别前,进行预处理的环节显得十分重要。本文对人脸图像的预处理工作主要有:彩色图像到灰度图像的转换、灰度归一化、几何归一化。几何归一化的主要工作有把检测到的大小不同的人脸图像的像素统一归一化为112×92,因为在研究人脸识别算法时,还会用到ORL人脸数据库,其中人脸图像像素大小正是112×92,为了最后在检验人脸识别算法时方便、快速,所以这里先对检测到的人脸图像进行几何归一化处理。
第三,人脸特征提取与识别,在这一章中,比较深入的研究分析了PCA、LDA人脸识别算法的原理,并对这几种算法的识别效果进行了验证,并且分析融合一种比较理想的人脸识别算法—PCA+LDA+BP,通过大量的实验表明该方法是有效可行的,一定程度上提高了人脸的识别率。
In recent years, with the rapid development of artificial intelligence and biotechnology, automatic face recognition has become a very active research topic in the image processing and pattern recognition.
In this paper, we study the basic theories and key technologies of face recognition. Then we mainly discuss the recognition problems in static color image in varying illumination. Main tasks are as follows:
Firstly, face detection. Face skin clustering color characteristic is studied in the paper. By the comparison of clustering results in various color space, we further improve the HSI color model and build the H-SI-I color model. Then we build face image with 100 face image database TXK and establish H-SI-I skin color model through a number of experiments. At last, simple detection, filtering and detailed examination to the face image in TXK are achieved in turn, and position face finally. A lot of experiments show that H-SI-I color model has a strong robustness for face detection of the various brightness of light.
Secondly, pre-processing. In order to improve the accuracy of face recognition, the face image preprocessing work is necessary before face recognition. Face image preprocessing work includes a lot of contents, such as filtering, gray-scale conversion, gray-scale normalization, geometry normalized and so on. Good pre-treatment will directly improve the ultimate rate of face recognition. Therefore, the task of pre-treatment is very important before a series of face recognition methods and work. In this paper, pre-treatment includes: the conversion from color image to grayscale image, gray normalization, geometric normalization, that is, the size of the detected face images normalized uniform into 112×92, because in the studying face recognition algorithms we will use ORL face image database, which is the image pixel size of 112×92. So the paper carries out geometric normalized of face image.
Thirdly, face feature extraction and recognition. Here we further analyze the principle of PCA, LDA recognition algorithms, and verify results of the algorithms. A face recognition Algorithm called PCA + LDA + BP is obtained by comparative analysis of integration. A large number of experiments show that the method is effective and practicable. To some extent, it improves the recognition rate of face.
本文在研究了人脸识别的基本理论和关键技术的基础上,重点讨论了不同光照条件下,彩色静止图像的人脸检测和识别问题。主要工作归纳如下:
第一,人脸检测。首先,本文对人脸肤色的聚类特性进行了研究,通过不同颜色空间中肤色聚类效果的比较,进一步改进了HSI颜色模型,构建了H-SI-I颜色模型。其次,自建了具有100张图片的人脸图像库TXK,并通过大量实验建立了H-SI-I肤色模型。最后,对人脸图像库TXK中的图像进行粗检、滤波、细检,最终定位出人脸。实验证明基于H-SI-I肤色模型的人脸检测算法是非常有效的。而且它对光照亮度也具有很强的鲁棒性。
第二,人脸图像预处理。为了提高人脸识别的精度,在人脸识别之前都需要进行人脸图像的预处理工作。预处理包括很多内容,例如对人脸图像进行滤波去噪、灰度转换、灰度归一化、几何归一化等。好的预处理将会直接提高最终的人脸识别率。因此在人脸识别前,进行预处理的环节显得十分重要。本文对人脸图像的预处理工作主要有:彩色图像到灰度图像的转换、灰度归一化、几何归一化。几何归一化的主要工作有把检测到的大小不同的人脸图像的像素统一归一化为112×92,因为在研究人脸识别算法时,还会用到ORL人脸数据库,其中人脸图像像素大小正是112×92,为了最后在检验人脸识别算法时方便、快速,所以这里先对检测到的人脸图像进行几何归一化处理。
第三,人脸特征提取与识别,在这一章中,比较深入的研究分析了PCA、LDA人脸识别算法的原理,并对这几种算法的识别效果进行了验证,并且分析融合一种比较理想的人脸识别算法—PCA+LDA+BP,通过大量的实验表明该方法是有效可行的,一定程度上提高了人脸的识别率。
In recent years, with the rapid development of artificial intelligence and biotechnology, automatic face recognition has become a very active research topic in the image processing and pattern recognition.
In this paper, we study the basic theories and key technologies of face recognition. Then we mainly discuss the recognition problems in static color image in varying illumination. Main tasks are as follows:
Firstly, face detection. Face skin clustering color characteristic is studied in the paper. By the comparison of clustering results in various color space, we further improve the HSI color model and build the H-SI-I color model. Then we build face image with 100 face image database TXK and establish H-SI-I skin color model through a number of experiments. At last, simple detection, filtering and detailed examination to the face image in TXK are achieved in turn, and position face finally. A lot of experiments show that H-SI-I color model has a strong robustness for face detection of the various brightness of light.
Secondly, pre-processing. In order to improve the accuracy of face recognition, the face image preprocessing work is necessary before face recognition. Face image preprocessing work includes a lot of contents, such as filtering, gray-scale conversion, gray-scale normalization, geometry normalized and so on. Good pre-treatment will directly improve the ultimate rate of face recognition. Therefore, the task of pre-treatment is very important before a series of face recognition methods and work. In this paper, pre-treatment includes: the conversion from color image to grayscale image, gray normalization, geometric normalization, that is, the size of the detected face images normalized uniform into 112×92, because in the studying face recognition algorithms we will use ORL face image database, which is the image pixel size of 112×92. So the paper carries out geometric normalized of face image.
Thirdly, face feature extraction and recognition. Here we further analyze the principle of PCA, LDA recognition algorithms, and verify results of the algorithms. A face recognition Algorithm called PCA + LDA + BP is obtained by comparative analysis of integration. A large number of experiments show that the method is effective and practicable. To some extent, it improves the recognition rate of face.
引文
[1]景英娟,董育宁.生物特征识别技术综述[J].桂林电子工业学院学报,2005,25(2):28-33.
[2]孙冬梅,裘正定.生物特征识别技术综述[J].电子学报,2001,29(12):1744-1748.
[3]肖冰,王映辉.人脸识别研究综述[J].计算机应用研究,2005,(8):1-5.
[4]李刚,高政,人脸自动识别方法综述[J].计算机应用研究,2003,20(8):4-9.
[5]段锦.人脸自动识别中若干问题的研究:(硕士学位论文).吉林:吉林大学,2004.
[6]赵丽红.人脸检测和识别算法的研究与实现[D].沈阳:东北大学,2006:8.
[7]丁海波,薛质,李生红.基于HIS空间的肤色检测方[J].计算机应用,2004,24:210-211.
[8]Beveridge,JR,Factors that influence algorithm periformace in the Face Recognition Grand Challenge[J].Computer Vision And Image Understanding 2009,6(113):750-762.
[9]陈传波,金先级.数字图像处理[M].机械工业出版,2004.
[10]刘文达,胡荣强等.基于肤色和模版匹配模型的人脸识别[J].中国高新技术企业,2008,24:144-145.
[11]郭红建,敖婷,冯建强.复杂背景彩色图像中的人脸分割技术[J].计算机工程与应用,2005,41(35):73-75.
[12]章毓晋.图象处理和分析[M].北京:清华大学出版社,1999.
[13]Park,S,Kim.D.Subtle facial expression recognition using motion magnifaction[J].Pattern Recognitionletters.2009,7(30):708-716.
[14]冈莎雷斯著.数字图像处理[M].阮秋琦,阮宇智译.北京:电子工业出版社,2003.
[15]贾永红,数字图象处理[M],武汉武汉大学出版社,2003,63-72,132-141.
[16]张翠平,苏光.人脸识别技术综述.中国图像图形学报,2000.10.
[17]何斌,马天予.Visual C++数字图像处理(第二版)[M].北京,人民邮电出版社,2002:199-200.
[18]庞彦伟等.融合奇异值分解和线性辨别分析的人脸识别算法[J].电路与系统学报.2006.8,11(4).
[19]何国辉,甘俊英.PCA+LDA算法在性别鉴别中的应用[J].计算机工程,2006.10, 30(19).
[20]陶亮.基于人脸识别的身份认证方法研究:(博士学位论文).合肥:中国科学技术大学,2003.
[21]许嘉伟,冯国灿等.DCT域上人脸光照正则化及识别[J].计算机应用,2009,29(1):47-50.
[22]Ekinci,M,Aykut,M.Palmprint recognition by applying wavelet-based kernel PCA [J].Journal of Computer Science and Technology.2008,(23):851-861.
[23]Zheng WS.Perturbation LDA:Learning the difference between the class empirical mean and its expectation[J].Pattern Recognition.2009,5(42):764-779.
[24]Kwak N.Feature extraction for one-class classification problems:Enhancements to biased discriminant analysis[J].Pattern Recognition..2009,1(42):17-26.
[25]李学桥.神经网络工程应用[M].重庆:重庆大学出版社,1995.
[26]王建国,杨万和,郑宇杰.一种基于ICA和模糊LDA的特征提取方法[J].模式识别和人工智能,2008,6(21):819-823.
[27]闻新等.MATLAB神经网络仿真与应用[M].科学出版社,2003.
[28]李晓东,张涛.一种改进的模块PCA方法及其在人脸识别中的应用[J].测试技术,2008,11(27):19-21.
[29]张熠,熊飞,张桂林.一种光照不变人脸识别的预处理算法[J].中国图象图形学报,2008,9:1707-1712.
[30]孙即祥等.现代模式识别[M].国防科技大学出版社,2002.1.
[31]周杰,卢春雨,张长水,李衍达.人脸自动识别方法综述[J].电子学报,2004(4):102-106.
[32]魏莱,王守觉,徐菲菲.近邻边界Fisher判别分析[J].电子与信息学报,2009,3:509-513.
[33]马平,靳敬永.改进的线性判别分析及人脸识别[J].计算机与数字工程,2009,1(37):135-137.
[34]Rafael C.Gonzalez Richad E.Woods.数字图像处理(第二版)[M].阮秋琦,阮宇智.北京.电于工业出版社,2003.
[35]何倩,冯国灿.奇异值分解在人脸识别中的应用[J].广东教育学院学报,2006,26(3).
[36]洪子泉,杨静宇.基于奇异值特征和统计模型的人像识别算法[J].计算机研究与 发展,1994,31(3).
[2]孙冬梅,裘正定.生物特征识别技术综述[J].电子学报,2001,29(12):1744-1748.
[3]肖冰,王映辉.人脸识别研究综述[J].计算机应用研究,2005,(8):1-5.
[4]李刚,高政,人脸自动识别方法综述[J].计算机应用研究,2003,20(8):4-9.
[5]段锦.人脸自动识别中若干问题的研究:(硕士学位论文).吉林:吉林大学,2004.
[6]赵丽红.人脸检测和识别算法的研究与实现[D].沈阳:东北大学,2006:8.
[7]丁海波,薛质,李生红.基于HIS空间的肤色检测方[J].计算机应用,2004,24:210-211.
[8]Beveridge,JR,Factors that influence algorithm periformace in the Face Recognition Grand Challenge[J].Computer Vision And Image Understanding 2009,6(113):750-762.
[9]陈传波,金先级.数字图像处理[M].机械工业出版,2004.
[10]刘文达,胡荣强等.基于肤色和模版匹配模型的人脸识别[J].中国高新技术企业,2008,24:144-145.
[11]郭红建,敖婷,冯建强.复杂背景彩色图像中的人脸分割技术[J].计算机工程与应用,2005,41(35):73-75.
[12]章毓晋.图象处理和分析[M].北京:清华大学出版社,1999.
[13]Park,S,Kim.D.Subtle facial expression recognition using motion magnifaction[J].Pattern Recognitionletters.2009,7(30):708-716.
[14]冈莎雷斯著.数字图像处理[M].阮秋琦,阮宇智译.北京:电子工业出版社,2003.
[15]贾永红,数字图象处理[M],武汉武汉大学出版社,2003,63-72,132-141.
[16]张翠平,苏光.人脸识别技术综述.中国图像图形学报,2000.10.
[17]何斌,马天予.Visual C++数字图像处理(第二版)[M].北京,人民邮电出版社,2002:199-200.
[18]庞彦伟等.融合奇异值分解和线性辨别分析的人脸识别算法[J].电路与系统学报.2006.8,11(4).
[19]何国辉,甘俊英.PCA+LDA算法在性别鉴别中的应用[J].计算机工程,2006.10, 30(19).
[20]陶亮.基于人脸识别的身份认证方法研究:(博士学位论文).合肥:中国科学技术大学,2003.
[21]许嘉伟,冯国灿等.DCT域上人脸光照正则化及识别[J].计算机应用,2009,29(1):47-50.
[22]Ekinci,M,Aykut,M.Palmprint recognition by applying wavelet-based kernel PCA [J].Journal of Computer Science and Technology.2008,(23):851-861.
[23]Zheng WS.Perturbation LDA:Learning the difference between the class empirical mean and its expectation[J].Pattern Recognition.2009,5(42):764-779.
[24]Kwak N.Feature extraction for one-class classification problems:Enhancements to biased discriminant analysis[J].Pattern Recognition..2009,1(42):17-26.
[25]李学桥.神经网络工程应用[M].重庆:重庆大学出版社,1995.
[26]王建国,杨万和,郑宇杰.一种基于ICA和模糊LDA的特征提取方法[J].模式识别和人工智能,2008,6(21):819-823.
[27]闻新等.MATLAB神经网络仿真与应用[M].科学出版社,2003.
[28]李晓东,张涛.一种改进的模块PCA方法及其在人脸识别中的应用[J].测试技术,2008,11(27):19-21.
[29]张熠,熊飞,张桂林.一种光照不变人脸识别的预处理算法[J].中国图象图形学报,2008,9:1707-1712.
[30]孙即祥等.现代模式识别[M].国防科技大学出版社,2002.1.
[31]周杰,卢春雨,张长水,李衍达.人脸自动识别方法综述[J].电子学报,2004(4):102-106.
[32]魏莱,王守觉,徐菲菲.近邻边界Fisher判别分析[J].电子与信息学报,2009,3:509-513.
[33]马平,靳敬永.改进的线性判别分析及人脸识别[J].计算机与数字工程,2009,1(37):135-137.
[34]Rafael C.Gonzalez Richad E.Woods.数字图像处理(第二版)[M].阮秋琦,阮宇智.北京.电于工业出版社,2003.
[35]何倩,冯国灿.奇异值分解在人脸识别中的应用[J].广东教育学院学报,2006,26(3).
[36]洪子泉,杨静宇.基于奇异值特征和统计模型的人像识别算法[J].计算机研究与 发展,1994,31(3).