基于自适应变换和空域同态变换的人脸光照补偿方法研究
|
摘要
人脸检测有重要的科研和应用价值,AdaBoost是当前执行检测速度最快的算法之一。光照问题是人脸检测和人脸识别研究的主要问题之一,是限制人脸检测率和人脸识别率提高的主要原因。光照补偿方法分成两个大类:基于图像处理的方法和基于模型的方法,基于模型的方法因计算复杂和限制条件较多应用并不广泛,基于图像处理的方法因计算简单速度快得到广泛应用。基于图像处理的方法在人脸检测中表现优秀的GIC、HE和HS等方法都存在一些问题,GIC需要先指定参照光照图像,HS需要指定直方图形状,HE容易受图像背景影响,而对数、指数和伽玛校正等方法光照处理能力不够强,对数变换只能补偿过暗偏光图像,指数变换只能补偿过亮图像,伽玛校正无法独立补偿图像光照。本文采用自适应变换技术和空域同态技术进现有常见方法进行改进,使基于图像处理的光照补偿方法的整体性能得到提高,提高了人脸检测率。
光照问题非常复杂,所有的基于图像处理方法的光照补偿对正确曝光图像都有一定的损害,直接导制人脸检测率的降低,造成这一问题的主要原因是因为没有一个行之有效的判断光照的标准,使无需光照补偿的图像和需要光照补偿的图像一起被光照补偿方法处理了。本文定义了二个简单的图像光照标准,建立在此二个光照标准上的自适应变换技术大大改善了以上情况,自适应对数变换和自适应指数变换可以对过暗和过亮光照图像进行有效补偿,自适应伽玛校正无需参数图像或变换参数可以自动对所有有问题光照进行光照补偿,采用自适应技术的直方图均匀化也在一定程度改善了受背景影响的缺点。
另外,基于图像处理的光照补偿方法没有区分光照和反射,会把图像的反射和图像光照一起处理,造成图像的失真,本文提出了空域同态补偿技术解决这个问题,同态补偿技术可以应用到对数变换、指数变换、伽玛变换等光照补偿方法中,使被补偿后的图像细节变得更清晰。
光照问题只是人脸检测和人脸识别研究中的一个方面,所以,在最后本文采用了磁盘蔟聚技术和异步调回I/O技术来提高AdaBoost的训练速度训练了一个1000个弱分类器的级联分类器,做为支撑前面提到的光照技术的研究和应用平台。
Face detection and face recognition is of important research and application value, and AdaBoost is one of the fastest algorithm for face detection. Illumination is an principal problem research and application for face detection and face recognition, it is main cause that constraint increase of face detection rate and face recognition rate. The methods of illumination compensation is divided into two categories: based-processing image and based-module, the based-module method is not widely application because its computation is complicate and its constraint condition is many, the based-processing image method is in widespread use because its computation is simply and fast. The based-processing image methods, HE、GIC and HS is of something in its, GIC need a corresponding standard illumination image, HE is easy to been influenced by background, HS need propose a histogram curve, apart from this, logarithm transform, exponent transform and gamma correction its abilities to process illumination compensation are not adequate, logarithm transform is only applicable to too dark image, exponent transform is only applicable to too bright image, gamma correction can not proform independently illumination compensation. In this paper propose adapted transform techique and airspace domain homomorphic transform techique that can been applied by common illumination compensation methods, its can improve proformance of face detection.
The illumination of image is very complicate, almost all based-processing image methods can injury normal illumination images, and reduce decrease of face detection, the main cause that lead to this problem is no criterion of illumination, and standard illumination images and those images that need compensate are all proformed. In this paper two simply illumination criterion was proposed, and adapted transform techique based these criterion greatly improve all of the above, adapted logarithm and adapted exponent can been applicable to dark image and bright image, adapted gamma correction proform without any parameter and compensate all illumination images, HE adopted adaptive transform techique has eliminated defection of easy to influented by background to some extend.
The based-processing image methods did not distinguish incoming and reflecting of image, and the incoming and the reflecting are to been process altogather, bring about distortion of image. In this paper the airspace homomorphic compensation techique was proposed to solve this problem, this techique can been widespeard used by logarithm transform, exponent transform and gamma correction etc., image that proformed will become more clarity in detail.
Illumination is only an aspect of the research of face detection and face recognition, in order to train a Adaboost system this paper adopted disk nest poly technology and asynchronous back I/O technology to increase the speed of training, and 1000 weak classifier was boosted into an strong classifer, this is to be platform of research and application of face detection and face recognition.
光照问题非常复杂,所有的基于图像处理方法的光照补偿对正确曝光图像都有一定的损害,直接导制人脸检测率的降低,造成这一问题的主要原因是因为没有一个行之有效的判断光照的标准,使无需光照补偿的图像和需要光照补偿的图像一起被光照补偿方法处理了。本文定义了二个简单的图像光照标准,建立在此二个光照标准上的自适应变换技术大大改善了以上情况,自适应对数变换和自适应指数变换可以对过暗和过亮光照图像进行有效补偿,自适应伽玛校正无需参数图像或变换参数可以自动对所有有问题光照进行光照补偿,采用自适应技术的直方图均匀化也在一定程度改善了受背景影响的缺点。
另外,基于图像处理的光照补偿方法没有区分光照和反射,会把图像的反射和图像光照一起处理,造成图像的失真,本文提出了空域同态补偿技术解决这个问题,同态补偿技术可以应用到对数变换、指数变换、伽玛变换等光照补偿方法中,使被补偿后的图像细节变得更清晰。
光照问题只是人脸检测和人脸识别研究中的一个方面,所以,在最后本文采用了磁盘蔟聚技术和异步调回I/O技术来提高AdaBoost的训练速度训练了一个1000个弱分类器的级联分类器,做为支撑前面提到的光照技术的研究和应用平台。
Face detection and face recognition is of important research and application value, and AdaBoost is one of the fastest algorithm for face detection. Illumination is an principal problem research and application for face detection and face recognition, it is main cause that constraint increase of face detection rate and face recognition rate. The methods of illumination compensation is divided into two categories: based-processing image and based-module, the based-module method is not widely application because its computation is complicate and its constraint condition is many, the based-processing image method is in widespread use because its computation is simply and fast. The based-processing image methods, HE、GIC and HS is of something in its, GIC need a corresponding standard illumination image, HE is easy to been influenced by background, HS need propose a histogram curve, apart from this, logarithm transform, exponent transform and gamma correction its abilities to process illumination compensation are not adequate, logarithm transform is only applicable to too dark image, exponent transform is only applicable to too bright image, gamma correction can not proform independently illumination compensation. In this paper propose adapted transform techique and airspace domain homomorphic transform techique that can been applied by common illumination compensation methods, its can improve proformance of face detection.
The illumination of image is very complicate, almost all based-processing image methods can injury normal illumination images, and reduce decrease of face detection, the main cause that lead to this problem is no criterion of illumination, and standard illumination images and those images that need compensate are all proformed. In this paper two simply illumination criterion was proposed, and adapted transform techique based these criterion greatly improve all of the above, adapted logarithm and adapted exponent can been applicable to dark image and bright image, adapted gamma correction proform without any parameter and compensate all illumination images, HE adopted adaptive transform techique has eliminated defection of easy to influented by background to some extend.
The based-processing image methods did not distinguish incoming and reflecting of image, and the incoming and the reflecting are to been process altogather, bring about distortion of image. In this paper the airspace homomorphic compensation techique was proposed to solve this problem, this techique can been widespeard used by logarithm transform, exponent transform and gamma correction etc., image that proformed will become more clarity in detail.
Illumination is only an aspect of the research of face detection and face recognition, in order to train a Adaboost system this paper adopted disk nest poly technology and asynchronous back I/O technology to increase the speed of training, and 1000 weak classifier was boosted into an strong classifer, this is to be platform of research and application of face detection and face recognition.
引文
[1] B Du, S.G Shan, L.Y. Qing et al., Empirical comparisons of several preprocessing methods for illumination insensitive face recognition, in: International Conference on Acoustics, Speech and Signal Processing, 2005:81–984.C.
[2] Shiguang Shan,Wen Gao,Bo Cao,Debin Zhao.Illumination Normalization for Robust Face RecognitionAgainst Varying Lighting Conditions[C] //Proc of IEEE International Workshop on Analysis and Modeling of Faces and Gestures. 2003 AMFG: 157-164.
[3]段锦.人脸自动机器识别[M].北京:科学出版社,2009.
[4]付忠良.关于AdaBoost有效性的分析[J].计算机研究和发展,2008,45(10):1747-1755.
[5] FERET. Face recognition vendor test 2002[R]. Evaluation Report, 2002.
[6]谭铁牛.中国生物特征认证动态[M].西安:西北大学出版社,2003.
[7] Chellappa R,Wilson C L, Sirohey S. Human and machine recognition of faces: a survey[J]. Proc. IEEE, 1995,83(5):705-740.
[8]周杰,卢春雨,张长水等.人脸自动识别方法综述[J].电子学报,2000,28(4):102~106.
[9]李刚,高政.人脸自动识别方法综述[J].计算机应用研究,2003,8:.
[10]周激流,张哗.人脸识别理论研究进展[J].计算机辅助设计与图形学学报,1999,11(2):180~184.
[11] PHILLIPS P J, MOON H, SYED A,et al.The FERET evaluationmethodology for face-recognition algorithms [J].IEEE TPAM ,I2000,22(10): 1090-1104.
[12] Moses Y,Ullman S.Limitation of Non-Model-Based Recognition Schemes[A].Proc ECCV-92,Sandini G,ed[C].Springer-Verlag,1992.
[13] Warrington E,Taylor A.Two Categorical Stages of Object Recogni-tion[J]. Perception, 1978, 7:152-164.
[14] Govindaraju V,Sher D B,Srihari R,Srihari S N.Locating Human Faces in Newspaper Photographs[A].Proc CVPP89[C],1989.549-554.
[15] MIAO J, YIN B C, WANG KQ,etal.A hierarchicalmultiscale and multiangle system for human face detection in a complex background using gravity-center template[J]. Pattern Recognition, 1999,32(6): 1237-1248.
[16] BUHMANN J, LADESM, BECKMAN F. Silicon retina object recog-nition,No 8596-CS[R]. Bonn: Institute for Informatics UniversityBonn, 1993
[17] Edelman S,Reisfeld D,Yeshurun Y.A System for Face Recognition that Learns from Examples[A].Proc,European Conf Computer Vision,Sandini S,ed[C].Springer-Verlag,1992:787-791.
[18] REISFELD D, YESHURUN Y. Robustdetection of facial features bygeneralized symmetry[C] //Proc of International Conference on Pat-tern Recognition. [S. .l ]: TheHagueNetherlands, 1992: 117-120.
[19]郑庆,闵帆,陈雷霆.基于复合变换的人脸照度补偿方案[J].计算机应用研究, 25(3.2),2008
[20]刘宏,李锦涛,苗军.多方法融合来解决人脸检测中的光照补偿.系统仿真学报[J],2001,13(Suppl):486-489.
[21] Reisfeld D,Yeshurun Y.Robust Detection of Facial Features by Generalized Symmetry[A].ICPR A,1992.117-120.
[22] ZHAOW, CHELLAPPA R. Illumination-insensitive face recognition using symmetric shape-from-shading[C] //Proc of IEEE Conference on CVPR. 2000: 286-293.
[23]卿来云,山世光,陈熙霖,等.基于球面谐波基图像的任意光照下的人脸识别[J].计算机学报, 2006,29(5): 760-768.
[24] HE Xiao-Guang, TIAN Jie, WU Li-Fang, ZHANG Yao-Yao, YANG Xin. Illumination Normalization with Morphological Quotient Image[J]. Journal of Software,2007,18(9):2318?2325.
[25] HE Xiao-Guang,TIAN Jie,WU Li-Fang,ZHANG Yao-Yao,YANG Xin.Illumination Normalization with Morphological Quotient Image, Journal of Software[J]. 18(9):2318-2325, 2007.9 http://www.jos.org.cn.
[26]卿来云,山世光,陈熙霖,等.基于球面谐波基图像的任意光照下的人脸识别[J].计算机学报, 2006,29(4): 760-76.
[27] K.Delac,M.Grgic,T.Kos, Sub-Image Homomorphic Filtering Technique for Improving Facial Identification under Difficult Illumination Conditions[C], International Conference on Systems,Signals and Image Processing(IWSSIP’06), September 21-23,2006.Budapest,Hungary
[28]张新明,沈兰荪.基于小波的同态滤波器用于图像对比度增强[J].电子学报,29(3.4):531-533,2001
[29]陈智,王国志,丰善,高宏文.同态滤波预处理在微粒场全息图图像处理中的应用[J].光子学报,33(3.2), pp.167-170,2004
[30]郑刚,贾振红.同态技术在红外图像处理中的应用[J],光子学报,34(9):1401-1403,2005
[31]刘宏,李锦涛,苗军.多方法融合来解决人脸检测中的照度补偿[J].系统仿真学报,13(增刊):486-489,2001
[32]王晓勇,胡福乔,赵宇明.同态滤波在图像角点提取中的应用[J].计算机工程, 32(11): 211-213, 2006
[33]冷建华.傅里叶变换[M].北京:清华大学出版社,2004:165~224.
[34]周琳娜,王东明,郭云彪,杨义先.基于数字图像边缘特性的形态学滤波取证技术[J].光子学报,36(8), pp.1047-1051,2008
[35]李梅,刘亦农.信息论基础教程[M].第2版.北京:北京邮电大学出版社,2008:7:24.
[36]王鸿南,钟文,汪静,夏德深.图像清晰度评价方法研究[J].中国图象图形学报,2004,9(7):828-831.
[37] Yang M H,Kriegman D,Ahuja N.Detecting faces in images:A survey[J].IEEE Trans.Pattern Analysis and Machine Intelligence,2002,24(3.1):34-58.
[38] Paul Viola,Michael J.Jones.Robust Real-Time Face Detection[J].Intemational Journal of Computer Vision 57(3.2),1 37-1 54,2004.
[39] Kapfer M,Benois-PineauJ. Detection of human faces in colorimage sequences with arbitrary motions for very low bitrate videophone coding[J].Pattern Recognition Letters,1997,18(14):1503~1518.
[40] Schapire R E. The strength of weak learnability. Machine Learning [J].1990,5(2):197~227.
[41] Valiant L G. A theory of the learnable [C]. Communications of the ACM,1984,27(11):1134~1142.
[42]蒋立辉,李宁,成向阳,王骐,尚铁梁.基于一种新的同态滤波散斑噪声压缩[J].激光和红外,30(3.1):11-14,2000.
[43]唐明,马颂德,肖京.基于模型的远红外序列图像自适应增强算法及其分析[J].计算机学报,23(8):893-896,2000
[44]王欣,王德隽.离散信号的滤波[M],北京:电子工业出版社,2002:128-141
[45]孙延奎.小波分析及其应用[M].北京:机械工业学版社,2005:1-281
[46]田村秀行编著,金喜子,乔双译.计算机图像处理[M].北京:科学出版社,2004:33-39
[47] Kenneth R.Castleman著,朱志刚,林学,石定机等译.数字图像处理[M].北京:电子工业出版社,2005:141-170,252-293
[48]张阿珍,刘政林,邹雪城,向祖权.基于双三次插值算法的图像缩放引擎的设计[J].微电子学与计算机,2007,24(1):49-51.
[49]冯慧君,陶素娟,李隆.基于双线性插值算法的图像放缩技术与实现[J].计算机应用与软件,2004,21(7):117-119.
[50] R G Keys.Cubic convolution interpolation for digital im-age processing.IEEE Transactions on Acoustics,Speech,and Signal Processing,1981,29(6):1153~1160.
[51] JOHN F.SOWA.Knowledge Representation[M].北京:机械工业出版社,2003.
[52] Paul Viola,Michael J.Jones.Robust Real-Time Face Detection[J].Intemational Journal of Computer Vision 57(3.2),137-154,2004.
[53] Graf H P, Cosatto E,Gibbon D,et al. Multimodal system for locating heads and faces[c]//Proceedings of the Second International Conference on Automatic Face and Gesture Recognition,1996,88~93
[2] Shiguang Shan,Wen Gao,Bo Cao,Debin Zhao.Illumination Normalization for Robust Face RecognitionAgainst Varying Lighting Conditions[C] //Proc of IEEE International Workshop on Analysis and Modeling of Faces and Gestures. 2003 AMFG: 157-164.
[3]段锦.人脸自动机器识别[M].北京:科学出版社,2009.
[4]付忠良.关于AdaBoost有效性的分析[J].计算机研究和发展,2008,45(10):1747-1755.
[5] FERET. Face recognition vendor test 2002[R]. Evaluation Report, 2002.
[6]谭铁牛.中国生物特征认证动态[M].西安:西北大学出版社,2003.
[7] Chellappa R,Wilson C L, Sirohey S. Human and machine recognition of faces: a survey[J]. Proc. IEEE, 1995,83(5):705-740.
[8]周杰,卢春雨,张长水等.人脸自动识别方法综述[J].电子学报,2000,28(4):102~106.
[9]李刚,高政.人脸自动识别方法综述[J].计算机应用研究,2003,8:.
[10]周激流,张哗.人脸识别理论研究进展[J].计算机辅助设计与图形学学报,1999,11(2):180~184.
[11] PHILLIPS P J, MOON H, SYED A,et al.The FERET evaluationmethodology for face-recognition algorithms [J].IEEE TPAM ,I2000,22(10): 1090-1104.
[12] Moses Y,Ullman S.Limitation of Non-Model-Based Recognition Schemes[A].Proc ECCV-92,Sandini G,ed[C].Springer-Verlag,1992.
[13] Warrington E,Taylor A.Two Categorical Stages of Object Recogni-tion[J]. Perception, 1978, 7:152-164.
[14] Govindaraju V,Sher D B,Srihari R,Srihari S N.Locating Human Faces in Newspaper Photographs[A].Proc CVPP89[C],1989.549-554.
[15] MIAO J, YIN B C, WANG KQ,etal.A hierarchicalmultiscale and multiangle system for human face detection in a complex background using gravity-center template[J]. Pattern Recognition, 1999,32(6): 1237-1248.
[16] BUHMANN J, LADESM, BECKMAN F. Silicon retina object recog-nition,No 8596-CS[R]. Bonn: Institute for Informatics UniversityBonn, 1993
[17] Edelman S,Reisfeld D,Yeshurun Y.A System for Face Recognition that Learns from Examples[A].Proc,European Conf Computer Vision,Sandini S,ed[C].Springer-Verlag,1992:787-791.
[18] REISFELD D, YESHURUN Y. Robustdetection of facial features bygeneralized symmetry[C] //Proc of International Conference on Pat-tern Recognition. [S. .l ]: TheHagueNetherlands, 1992: 117-120.
[19]郑庆,闵帆,陈雷霆.基于复合变换的人脸照度补偿方案[J].计算机应用研究, 25(3.2),2008
[20]刘宏,李锦涛,苗军.多方法融合来解决人脸检测中的光照补偿.系统仿真学报[J],2001,13(Suppl):486-489.
[21] Reisfeld D,Yeshurun Y.Robust Detection of Facial Features by Generalized Symmetry[A].ICPR A,1992.117-120.
[22] ZHAOW, CHELLAPPA R. Illumination-insensitive face recognition using symmetric shape-from-shading[C] //Proc of IEEE Conference on CVPR. 2000: 286-293.
[23]卿来云,山世光,陈熙霖,等.基于球面谐波基图像的任意光照下的人脸识别[J].计算机学报, 2006,29(5): 760-768.
[24] HE Xiao-Guang, TIAN Jie, WU Li-Fang, ZHANG Yao-Yao, YANG Xin. Illumination Normalization with Morphological Quotient Image[J]. Journal of Software,2007,18(9):2318?2325.
[25] HE Xiao-Guang,TIAN Jie,WU Li-Fang,ZHANG Yao-Yao,YANG Xin.Illumination Normalization with Morphological Quotient Image, Journal of Software[J]. 18(9):2318-2325, 2007.9 http://www.jos.org.cn.
[26]卿来云,山世光,陈熙霖,等.基于球面谐波基图像的任意光照下的人脸识别[J].计算机学报, 2006,29(4): 760-76.
[27] K.Delac,M.Grgic,T.Kos, Sub-Image Homomorphic Filtering Technique for Improving Facial Identification under Difficult Illumination Conditions[C], International Conference on Systems,Signals and Image Processing(IWSSIP’06), September 21-23,2006.Budapest,Hungary
[28]张新明,沈兰荪.基于小波的同态滤波器用于图像对比度增强[J].电子学报,29(3.4):531-533,2001
[29]陈智,王国志,丰善,高宏文.同态滤波预处理在微粒场全息图图像处理中的应用[J].光子学报,33(3.2), pp.167-170,2004
[30]郑刚,贾振红.同态技术在红外图像处理中的应用[J],光子学报,34(9):1401-1403,2005
[31]刘宏,李锦涛,苗军.多方法融合来解决人脸检测中的照度补偿[J].系统仿真学报,13(增刊):486-489,2001
[32]王晓勇,胡福乔,赵宇明.同态滤波在图像角点提取中的应用[J].计算机工程, 32(11): 211-213, 2006
[33]冷建华.傅里叶变换[M].北京:清华大学出版社,2004:165~224.
[34]周琳娜,王东明,郭云彪,杨义先.基于数字图像边缘特性的形态学滤波取证技术[J].光子学报,36(8), pp.1047-1051,2008
[35]李梅,刘亦农.信息论基础教程[M].第2版.北京:北京邮电大学出版社,2008:7:24.
[36]王鸿南,钟文,汪静,夏德深.图像清晰度评价方法研究[J].中国图象图形学报,2004,9(7):828-831.
[37] Yang M H,Kriegman D,Ahuja N.Detecting faces in images:A survey[J].IEEE Trans.Pattern Analysis and Machine Intelligence,2002,24(3.1):34-58.
[38] Paul Viola,Michael J.Jones.Robust Real-Time Face Detection[J].Intemational Journal of Computer Vision 57(3.2),1 37-1 54,2004.
[39] Kapfer M,Benois-PineauJ. Detection of human faces in colorimage sequences with arbitrary motions for very low bitrate videophone coding[J].Pattern Recognition Letters,1997,18(14):1503~1518.
[40] Schapire R E. The strength of weak learnability. Machine Learning [J].1990,5(2):197~227.
[41] Valiant L G. A theory of the learnable [C]. Communications of the ACM,1984,27(11):1134~1142.
[42]蒋立辉,李宁,成向阳,王骐,尚铁梁.基于一种新的同态滤波散斑噪声压缩[J].激光和红外,30(3.1):11-14,2000.
[43]唐明,马颂德,肖京.基于模型的远红外序列图像自适应增强算法及其分析[J].计算机学报,23(8):893-896,2000
[44]王欣,王德隽.离散信号的滤波[M],北京:电子工业出版社,2002:128-141
[45]孙延奎.小波分析及其应用[M].北京:机械工业学版社,2005:1-281
[46]田村秀行编著,金喜子,乔双译.计算机图像处理[M].北京:科学出版社,2004:33-39
[47] Kenneth R.Castleman著,朱志刚,林学,石定机等译.数字图像处理[M].北京:电子工业出版社,2005:141-170,252-293
[48]张阿珍,刘政林,邹雪城,向祖权.基于双三次插值算法的图像缩放引擎的设计[J].微电子学与计算机,2007,24(1):49-51.
[49]冯慧君,陶素娟,李隆.基于双线性插值算法的图像放缩技术与实现[J].计算机应用与软件,2004,21(7):117-119.
[50] R G Keys.Cubic convolution interpolation for digital im-age processing.IEEE Transactions on Acoustics,Speech,and Signal Processing,1981,29(6):1153~1160.
[51] JOHN F.SOWA.Knowledge Representation[M].北京:机械工业出版社,2003.
[52] Paul Viola,Michael J.Jones.Robust Real-Time Face Detection[J].Intemational Journal of Computer Vision 57(3.2),137-154,2004.
[53] Graf H P, Cosatto E,Gibbon D,et al. Multimodal system for locating heads and faces[c]//Proceedings of the Second International Conference on Automatic Face and Gesture Recognition,1996,88~93