基于语义稀疏表示的不良图像检测算法
|
摘要
高质量的多媒体通信使人们易于获取大量的有用信息。然而,淫秽、色情等不良信息的传播方式也更为隐蔽化和多样化,而且带来严重的社会问题。因此该方面的研究成果对净化社会环境、保护青少年健康成长,维护社会伦理道德等方面有积极的作用。本文系统分析了现有方法的优缺点,针对如下问题进行较深入的研究:鲁棒的肤色建模、不良特征表示和判别。主要研究成果如下:
针对彩色不良图像中包含裸露的皮肤区域的特点,提出一种基于主动学习的贝叶斯肤色建模方法。根据YCbCr色彩空间与人类视觉感知机理相类似的特点和该空间中肤色的聚类特性,选择在YCbCr色彩空间上建立贝叶斯肤色判别模型;进而,通过Bootstrap主动反馈方法提升肤色训练样本的典型性和多样性,避免肤色模型发生过拟合。
为在肤色区域内引入高层语义信息,提高不良图像判别的准确度,本文在肤色区域内检测高斯差分算子的极值点得到图像的感兴趣点,根据SIFT描述子对平移、旋转、尺度缩放、亮度变化、遮挡和噪声等具有良好的不变性,对仿射变换也保持一定程度的稳定性,本文采用SIFT算法对兴趣点进行描述,提取出能更好体现这类图像的特征。
考虑到待测图像的多样性,本文提出基于词袋模型和稀疏表示的不良信息判别方法。SIFT特征匹配较为敏感,于是本文借鉴词袋模型将图像看作由码本中的码词组合而成。在不同的距离侧度下,通过K-均值聚类降低特征的敏感性,为不良图像和正常图像分别建立码本。将测试图像在两类码本分别进行稀疏表示,通过两类图像的重构误差来判断待测图像的性质。
论文工作从实际应用中抽象出科学问题,涉及到计算机视觉和统计机器学习的最新理论。研究内容富有前瞻性和挑战性,具有极其重要的理论意义和应用价值。
People can get plenty of useful information through high-quality multimedia communications. However, the transmission of pornographic and other undesirable information is also more subtle and concealed. This trend leads to serious social problems. Therefore, the research achievements for this purpose are helpful to purify the social environment, protect the mental health of the underages, and safeguard social ethics. The existing methods are systemically reviewed in this paper. Some crucial problems are discussed, such as, robust skin modeling, feature representation and evaluation of pornographic images. The main achievements of this paper are summarized as follows.
Since pornographic images always contain skin area, we proposed an active learning based Bayes discriminative model for skin detection. According to the similar mechanism of perception between the human visual system and YCbCr color space, and the existence of clustering trend in YCbCr color space, we choose the YCbCr color space to build our Bayes discriminative skin model. We further incorporate the Bootstrap method to improve the diversity and typicality of training samples to avoid model overfitting.
To improve the accuracy of image determination, we introduce the high level semantic information to the skin area analysis. The difference of Gaussian based extreme points detection operator is used to extract the is invariant to sifting, rotation, scale, brightness, occlusion and noise changes, and can maintain the stability for translation affine in certain degree. We adopt SIFT descriptor to represent the interesting points of the images, which can reflect the characteristics of the adult images better.
To handle the diversity of the test image, we propose a bag of words model and sparse representation based method for adult image evaluation. SIFT features are sensitive to points matching. We regard the images as the combination of the visual words, which come from the codebook. Under different distance measurements, the SIFT features are clustered by K-means clustering algorithm to reduce the feature sensitivity, through which the codebooks of normal images and adult images are formed, respectively. Then we represent the test images on these two codebooks by two reconstruction errors.
This work abstracts scientific issues from the practical applications. It involves the latest theories in computer vision and statistical machine learning areas. This research is forward-looking and challenging, it has an extremely important theoretical and practical value.
针对彩色不良图像中包含裸露的皮肤区域的特点,提出一种基于主动学习的贝叶斯肤色建模方法。根据YCbCr色彩空间与人类视觉感知机理相类似的特点和该空间中肤色的聚类特性,选择在YCbCr色彩空间上建立贝叶斯肤色判别模型;进而,通过Bootstrap主动反馈方法提升肤色训练样本的典型性和多样性,避免肤色模型发生过拟合。
为在肤色区域内引入高层语义信息,提高不良图像判别的准确度,本文在肤色区域内检测高斯差分算子的极值点得到图像的感兴趣点,根据SIFT描述子对平移、旋转、尺度缩放、亮度变化、遮挡和噪声等具有良好的不变性,对仿射变换也保持一定程度的稳定性,本文采用SIFT算法对兴趣点进行描述,提取出能更好体现这类图像的特征。
考虑到待测图像的多样性,本文提出基于词袋模型和稀疏表示的不良信息判别方法。SIFT特征匹配较为敏感,于是本文借鉴词袋模型将图像看作由码本中的码词组合而成。在不同的距离侧度下,通过K-均值聚类降低特征的敏感性,为不良图像和正常图像分别建立码本。将测试图像在两类码本分别进行稀疏表示,通过两类图像的重构误差来判断待测图像的性质。
论文工作从实际应用中抽象出科学问题,涉及到计算机视觉和统计机器学习的最新理论。研究内容富有前瞻性和挑战性,具有极其重要的理论意义和应用价值。
People can get plenty of useful information through high-quality multimedia communications. However, the transmission of pornographic and other undesirable information is also more subtle and concealed. This trend leads to serious social problems. Therefore, the research achievements for this purpose are helpful to purify the social environment, protect the mental health of the underages, and safeguard social ethics. The existing methods are systemically reviewed in this paper. Some crucial problems are discussed, such as, robust skin modeling, feature representation and evaluation of pornographic images. The main achievements of this paper are summarized as follows.
Since pornographic images always contain skin area, we proposed an active learning based Bayes discriminative model for skin detection. According to the similar mechanism of perception between the human visual system and YCbCr color space, and the existence of clustering trend in YCbCr color space, we choose the YCbCr color space to build our Bayes discriminative skin model. We further incorporate the Bootstrap method to improve the diversity and typicality of training samples to avoid model overfitting.
To improve the accuracy of image determination, we introduce the high level semantic information to the skin area analysis. The difference of Gaussian based extreme points detection operator is used to extract the is invariant to sifting, rotation, scale, brightness, occlusion and noise changes, and can maintain the stability for translation affine in certain degree. We adopt SIFT descriptor to represent the interesting points of the images, which can reflect the characteristics of the adult images better.
To handle the diversity of the test image, we propose a bag of words model and sparse representation based method for adult image evaluation. SIFT features are sensitive to points matching. We regard the images as the combination of the visual words, which come from the codebook. Under different distance measurements, the SIFT features are clustered by K-means clustering algorithm to reduce the feature sensitivity, through which the codebooks of normal images and adult images are formed, respectively. Then we represent the test images on these two codebooks by two reconstruction errors.
This work abstracts scientific issues from the practical applications. It involves the latest theories in computer vision and statistical machine learning areas. This research is forward-looking and challenging, it has an extremely important theoretical and practical value.
引文
[1] Fleck M, Forsyth D, Bregler C. Finding naked people. Proceedings of the 4th European Conference on Computer Vision. London: Springer-Verlag, pp.593-602, 1996.
[2] Jones M J, Rehg J M. Statistical color models with application to skin detection. International Journal of Computer Vision, 46(1): 81-96, 2002.
[3] Xu Xinxin, Yuan Hua. Pornographic image filtering based on Self-adaptive wavelet invariant regulation Method. Journal of Tsinghua University, Science and Technology. 45(1): 1878-1881, 2005.[徐欣欣,袁华.一种基于自适应小波不变矩的色情图像过滤方法.清华大学学报:自然科学版, 45(1): 1878-1881, 2005.]
[4] Zhao Xiaohui, Zhang Fuwei, Shen Xuanjing. Color skin algorithms and application based on Self-adaptive color space model. Journal of Jilin University, Information Science. 24(1):68-75, 2006.[赵晓晖,张福威,申铉京.基于亮度自适应色度空间模型肤色算法及应用.吉林大学学报:信息科学版, 24 (1) : 68-75, 2006.]
[5] Hu Guanyu. The study on naked people image detection based on skin color. Tainan: National Cheng Kung University,2004.[胡冠宇.基于肤色之裸体影像侦测之研究.台南:国立成功大学, 2004.]
[6] Yang Jinfeng, Fu Zhouyu, Tan Qieniu, etc. A new content-based image recognition and filtering method. Journal of China Institute of Communications, 25 (7): 93-106, 2004. [杨金锋,傅周宇,谭铁牛,等.一种新型的基于内容的图像识别与过滤方法.通信学报, 25 (7) : 93-106, 2004.]
[7] Drimbarean A F, Corcoran P M, Cuic M, et al. Image processing techniques to detect and filter objectionable images based on skin tone and shape recognition Proceedings of International Conference on Consumer Electronics. 2001: 278-279.
[8] Jeong C Y, Kim J S, Hong K S. Appearance-based nude image detection. Proceedings of the 17th International Conference on Pattern Recognition. Washington DC: IEEE Computer Society, pp.467-470, 2004.
[9] Peng Yuzhong, Yuan Changan, Wang Yan, Tan Xiao. Content-based understanding of harmful information filtering technology. Application Research Of Computers, 26(2): 433-438,2009. [彭昱忠,元昌安,王艳,覃晓.基于内容理解的不良信息过滤技术研究.计算机应用研究, 26(2): 433-438,2009.]
[10] Zheng H, Liu H, Daoudi M. Blocking objectionable images: Adult images and harmful symbols. ICME, pp.1223-1226, Taipei, Taiwan, 2004.
[11] Yoo J. Intelligent multimedia information retrieval for identifying and rating adult images. International Conference on KES, vol.1, pp.164-170, Wellington, NZ, 2004.
[12] Shih L, Lee H, Yang S. An adult image identification system employing image retrieval technique. Pattern Recognition Letters, 28(16):2367-2374, 2007.
[13] Roth D, Yang J, Jiang G, Hauptmann G, Ngo W. Evaluating bag-of-visual-wordsrepresentations in scene classification. ACM MIR, pp.197-206, NY, USA, 2007.
[14] Cao L, Li F. Spatially coherent latent topic model for concurrent segmentation and classification of objects and scenes. ICCV, Rio de Janeiro, Brazil, 2007.
[15] Li F, Fergus R, Torralba A. Recognizing and learning object categories. CVPR, Minnesota, USA, 2007.
[16] Qiu G. Indexing chromatic and achromatic patterns for content-based colour image retrieval. Pattern Recognition, 35:1675-1686, 2002.
[17] Deselaers T, Pimenidis L, Ney H. Bag-of-visual-words models for adult image classification and filtering. ICPR, Florida, USA, 2008.
[18] Ana L, Sandra A. A bag-of-features approach based on hue-sift descriptor for nude detection, ESPC, 24-28, Glasgow, Scotland, 2009.
[19] Rainer L, Rudolf H. Filtering adult image content with topic models. NY, USA, 2009.
[20] Crowley L, Bedrune M. Integration and control of reactive visual processes. ECCV. 2:47-58, Stockholm, Sweden, 1994.
[21] Saxe D, Foulds R. Toward robust skin identification in video images. ICAFGR, pp.379-384, Vermont, USA, 1996.
[22] Kjeldsen R, Kender J. Finding skin in color images. ICAFGR, pp.312-317, Vermont, USA, 1996.
[23] Chai D, Bouzerdoum A. A bayesian approach to skin color classification in YCbCr color space. IEEE Region Ten Conference. 2:421-424, Kuala Lumpur, Malaysia, 2000.
[24] Engelbrecht A P, Cloete I. Incremental learning using sensitivity analysis. IJCNN’99, International Joint Conference on Neural Networks. vol.2. pp.1350-1355, 1999.
[25] David D L, Catlett J. Heterogeneous uncertainty sampling for supervised learning. In Proceedings of the 11th International Conference on Machine Learning, pp.48-156, New Brunswick, NJ, 1994.
[26] Seung H S, Opper M, Sompolinsky H. Query by Committee. Proceedings of the 15th Annual ACM Workshop on Computational Learning Theory, 287-294, California, 1992.
[27] Freund Y, Seung H S, Samir E. Selective Sampling Using the Query by Committee Algorithm. Machine Learning, 28(23):133-168, 1997.
[28] Roy N, McCallum A. Toward optimal active learning through sampling estimation of error reduction. Proceedings of the 18th Machine Learning. 441-448, San Francisco, CA, 2001.
[29] Tong S, Koller D. Active learning for parameter estimation in bayesian networks. Proceedings of Advances in Neural Information Processing Systems, 647-653 Cambridge, MA: MIT Press, 2000.
[30] Lindenbaum M, Markovitch S, Rusakov D. Selective sampling for nearest neighbor classifiers, 54(2):125-152, 2004.
[31] Efron B. Bootstrap methods: another look at the Jackknife. Annals of Statistics, vol.7, 1979.
[32] Chunna T, Xinbo G, Li Jie. An example selection method for active learning based onembedded bootstrap algorithm. Journal of Computer Research and Development. 43(10):1706-1712, 2006. [田春娜,高新波,李洁.基于嵌入式Bootstrap的主动学习示例选择方法.计算机研究与发展, 2006, 43(10):1706-1712.]
[33] Lowe D. G. Object recognition from local scale-invariant features. Proc. International Conference on Computer Vision. Corfu, pp:1150-1157,Greece, 1999.
[34] Lowe D. G. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision. pp60 (2): 91-1102004.
[35] Feifei L, Perona P. A bayesian hierarchical model for learning natural scene categories. Proceedings of IEEE Computer Vision and Pattern Recognition, pp.524-531, 2005.
[36] Feifei L, Fergus R, Torralba A. Recognizing and learing object categories, CVPR, short course, 2007.
[37] Vogel J, Schiele B. On performance characterization and optimization for image retrieval. Pattern Recognition, vol.35, pp.1675-1686, 2002.
[38] Harris C, Stephens M. A combined corner and edge detector. Proceedings of the 4th Alvey Vision Conference. pp.147-151, 1988.
[39] Vidal-Naquet M, Ullman S. Object recognition with informative features and linear classification. Proceedings of IEEE International Conference on Computer Vision. pp.281-288, 2003.
[40] Maree R, Geurts P, Piater J, Wehenkel L. Random subwindows for robust image classification. Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pp.34-40, 2005.
[41] Barnard K, Duygulu P, Forsyth F, Blei D, Jordan M. Matching words and pictures. Journal of Machine Learning Research, pp.3:1107-1135, 2003.
[42] Bruno A, Olshausen B A, Field D J. Emergence of simple-cell receptive field properties by learning a sparse code for natural image. Nature, pp.381:607-609, 1996.
[43] Candes E J. Ridgelets: theory and applications. Ph.D dissertation, Stanford University, Stanford, CA, 1998.
[44] Pennec E L, Mallat S. Sparse geometric image representations with bandelets. IEEE Transaction on Image Processing, pp.14(4):423-438, 2005.
[45] Do M N, Vetterli M. The contourlet transform: an efficient directional multiresolution image representation. IEEE Transaction on Image Processing,pp.14 (12):2091-2106, 2005.
[46] Tibshirani R. Regress ion shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B. pp.58 (1): 267-288, 1996.
[47] Yan L, Xuanjing Sh,Debing Zh. Skin detection based on texture. Computer Engineering and Applications, pp.19: 74-77, 2003. [李雁,申铉京,赵德斌.基于纹理的皮肤检测.计算机工程与应用,pp.19:74-77, 2003.]
[2] Jones M J, Rehg J M. Statistical color models with application to skin detection. International Journal of Computer Vision, 46(1): 81-96, 2002.
[3] Xu Xinxin, Yuan Hua. Pornographic image filtering based on Self-adaptive wavelet invariant regulation Method. Journal of Tsinghua University, Science and Technology. 45(1): 1878-1881, 2005.[徐欣欣,袁华.一种基于自适应小波不变矩的色情图像过滤方法.清华大学学报:自然科学版, 45(1): 1878-1881, 2005.]
[4] Zhao Xiaohui, Zhang Fuwei, Shen Xuanjing. Color skin algorithms and application based on Self-adaptive color space model. Journal of Jilin University, Information Science. 24(1):68-75, 2006.[赵晓晖,张福威,申铉京.基于亮度自适应色度空间模型肤色算法及应用.吉林大学学报:信息科学版, 24 (1) : 68-75, 2006.]
[5] Hu Guanyu. The study on naked people image detection based on skin color. Tainan: National Cheng Kung University,2004.[胡冠宇.基于肤色之裸体影像侦测之研究.台南:国立成功大学, 2004.]
[6] Yang Jinfeng, Fu Zhouyu, Tan Qieniu, etc. A new content-based image recognition and filtering method. Journal of China Institute of Communications, 25 (7): 93-106, 2004. [杨金锋,傅周宇,谭铁牛,等.一种新型的基于内容的图像识别与过滤方法.通信学报, 25 (7) : 93-106, 2004.]
[7] Drimbarean A F, Corcoran P M, Cuic M, et al. Image processing techniques to detect and filter objectionable images based on skin tone and shape recognition Proceedings of International Conference on Consumer Electronics. 2001: 278-279.
[8] Jeong C Y, Kim J S, Hong K S. Appearance-based nude image detection. Proceedings of the 17th International Conference on Pattern Recognition. Washington DC: IEEE Computer Society, pp.467-470, 2004.
[9] Peng Yuzhong, Yuan Changan, Wang Yan, Tan Xiao. Content-based understanding of harmful information filtering technology. Application Research Of Computers, 26(2): 433-438,2009. [彭昱忠,元昌安,王艳,覃晓.基于内容理解的不良信息过滤技术研究.计算机应用研究, 26(2): 433-438,2009.]
[10] Zheng H, Liu H, Daoudi M. Blocking objectionable images: Adult images and harmful symbols. ICME, pp.1223-1226, Taipei, Taiwan, 2004.
[11] Yoo J. Intelligent multimedia information retrieval for identifying and rating adult images. International Conference on KES, vol.1, pp.164-170, Wellington, NZ, 2004.
[12] Shih L, Lee H, Yang S. An adult image identification system employing image retrieval technique. Pattern Recognition Letters, 28(16):2367-2374, 2007.
[13] Roth D, Yang J, Jiang G, Hauptmann G, Ngo W. Evaluating bag-of-visual-wordsrepresentations in scene classification. ACM MIR, pp.197-206, NY, USA, 2007.
[14] Cao L, Li F. Spatially coherent latent topic model for concurrent segmentation and classification of objects and scenes. ICCV, Rio de Janeiro, Brazil, 2007.
[15] Li F, Fergus R, Torralba A. Recognizing and learning object categories. CVPR, Minnesota, USA, 2007.
[16] Qiu G. Indexing chromatic and achromatic patterns for content-based colour image retrieval. Pattern Recognition, 35:1675-1686, 2002.
[17] Deselaers T, Pimenidis L, Ney H. Bag-of-visual-words models for adult image classification and filtering. ICPR, Florida, USA, 2008.
[18] Ana L, Sandra A. A bag-of-features approach based on hue-sift descriptor for nude detection, ESPC, 24-28, Glasgow, Scotland, 2009.
[19] Rainer L, Rudolf H. Filtering adult image content with topic models. NY, USA, 2009.
[20] Crowley L, Bedrune M. Integration and control of reactive visual processes. ECCV. 2:47-58, Stockholm, Sweden, 1994.
[21] Saxe D, Foulds R. Toward robust skin identification in video images. ICAFGR, pp.379-384, Vermont, USA, 1996.
[22] Kjeldsen R, Kender J. Finding skin in color images. ICAFGR, pp.312-317, Vermont, USA, 1996.
[23] Chai D, Bouzerdoum A. A bayesian approach to skin color classification in YCbCr color space. IEEE Region Ten Conference. 2:421-424, Kuala Lumpur, Malaysia, 2000.
[24] Engelbrecht A P, Cloete I. Incremental learning using sensitivity analysis. IJCNN’99, International Joint Conference on Neural Networks. vol.2. pp.1350-1355, 1999.
[25] David D L, Catlett J. Heterogeneous uncertainty sampling for supervised learning. In Proceedings of the 11th International Conference on Machine Learning, pp.48-156, New Brunswick, NJ, 1994.
[26] Seung H S, Opper M, Sompolinsky H. Query by Committee. Proceedings of the 15th Annual ACM Workshop on Computational Learning Theory, 287-294, California, 1992.
[27] Freund Y, Seung H S, Samir E. Selective Sampling Using the Query by Committee Algorithm. Machine Learning, 28(23):133-168, 1997.
[28] Roy N, McCallum A. Toward optimal active learning through sampling estimation of error reduction. Proceedings of the 18th Machine Learning. 441-448, San Francisco, CA, 2001.
[29] Tong S, Koller D. Active learning for parameter estimation in bayesian networks. Proceedings of Advances in Neural Information Processing Systems, 647-653 Cambridge, MA: MIT Press, 2000.
[30] Lindenbaum M, Markovitch S, Rusakov D. Selective sampling for nearest neighbor classifiers, 54(2):125-152, 2004.
[31] Efron B. Bootstrap methods: another look at the Jackknife. Annals of Statistics, vol.7, 1979.
[32] Chunna T, Xinbo G, Li Jie. An example selection method for active learning based onembedded bootstrap algorithm. Journal of Computer Research and Development. 43(10):1706-1712, 2006. [田春娜,高新波,李洁.基于嵌入式Bootstrap的主动学习示例选择方法.计算机研究与发展, 2006, 43(10):1706-1712.]
[33] Lowe D. G. Object recognition from local scale-invariant features. Proc. International Conference on Computer Vision. Corfu, pp:1150-1157,Greece, 1999.
[34] Lowe D. G. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision. pp60 (2): 91-1102004.
[35] Feifei L, Perona P. A bayesian hierarchical model for learning natural scene categories. Proceedings of IEEE Computer Vision and Pattern Recognition, pp.524-531, 2005.
[36] Feifei L, Fergus R, Torralba A. Recognizing and learing object categories, CVPR, short course, 2007.
[37] Vogel J, Schiele B. On performance characterization and optimization for image retrieval. Pattern Recognition, vol.35, pp.1675-1686, 2002.
[38] Harris C, Stephens M. A combined corner and edge detector. Proceedings of the 4th Alvey Vision Conference. pp.147-151, 1988.
[39] Vidal-Naquet M, Ullman S. Object recognition with informative features and linear classification. Proceedings of IEEE International Conference on Computer Vision. pp.281-288, 2003.
[40] Maree R, Geurts P, Piater J, Wehenkel L. Random subwindows for robust image classification. Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pp.34-40, 2005.
[41] Barnard K, Duygulu P, Forsyth F, Blei D, Jordan M. Matching words and pictures. Journal of Machine Learning Research, pp.3:1107-1135, 2003.
[42] Bruno A, Olshausen B A, Field D J. Emergence of simple-cell receptive field properties by learning a sparse code for natural image. Nature, pp.381:607-609, 1996.
[43] Candes E J. Ridgelets: theory and applications. Ph.D dissertation, Stanford University, Stanford, CA, 1998.
[44] Pennec E L, Mallat S. Sparse geometric image representations with bandelets. IEEE Transaction on Image Processing, pp.14(4):423-438, 2005.
[45] Do M N, Vetterli M. The contourlet transform: an efficient directional multiresolution image representation. IEEE Transaction on Image Processing,pp.14 (12):2091-2106, 2005.
[46] Tibshirani R. Regress ion shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B. pp.58 (1): 267-288, 1996.
[47] Yan L, Xuanjing Sh,Debing Zh. Skin detection based on texture. Computer Engineering and Applications, pp.19: 74-77, 2003. [李雁,申铉京,赵德斌.基于纹理的皮肤检测.计算机工程与应用,pp.19:74-77, 2003.]