基於相似群集的臉部偵測與辨識之研究

A Study for Face Detection and Recognition Based on A Similarity-based Clustering Algorithm

陳予郎(Yui-Lang Chen)

指導教授 ： 林文暉

崑山科技大學/資訊科技學院/資訊管理研究所/碩士(2006年)

摘要

在本論文中我們提供了一個基於相似群集的臉部偵測與辨識之研究。在臉部辨識方面，我們提出了兩種臉部偵測演算法之應用－單一臉人與多重人臉偵測。關於單一臉人的偵測方法，我們提出一基於Log-Sigmiod函數之動態門檻函數的最適臉部區域探勘演算法，並結合Moghaddam貝氏分析法為改善傳統臉部偵測法在運算處理複雜和降低雜訊方面為法突破的缺失。關於多重人臉的偵測方法，演算法主要藉由影像像素之位址和顏色特徵，運用SCM相似群集演算法以有效將影像中同質區域粗略的區隔開來，接著以我們所提出的區域合併與區域分割演算法使人臉區域能完整切割出來，經由初步人臉可能性區域篩選後再以Spectral Histogram方式擷取人臉可能性區之紋理特徵，運用SVM之類神經演算法以訓練和辨識真實人臉區域，此方法主要有效突破一般傳統方法以固定大小的遮罩無法偵測人臉大小多樣化的問題。
在人臉辨識方面，本論文主要提供一低複雜度、無監督式自我學習與超高速比對之人臉辨識系統。在低複雜度方法，我們以2D小波分析擷取次頻帶影像的位置係數變異量為特徵並組成一個特徵向量來代表一張影像，藉此以有效達到低複雜度的運算處理。在無監督式自我學習方面，我們使用SCM演算法架構系統自我組織學習的強大能力。在高速比對方面，我們提出有效的二段式比對的方法，基於SCM的人臉輪廓群集並藉由FLD與PCA模組化臉部輪廓，運用貝氏的方析以比對測試人臉所屬之臉部模組，再於所屬臉部模組中以NLSE的最小誤差法進行逐步的一對一比對，以此系統可在一龐大資料庫下達到超高速度的人臉影像辨識。

關鍵字

小波分析 ； 主成份分析 ； 相似群集法 ； 貝氏分析

並列摘要

In this thesis, we propose a study for face detection and recognition based on similarity clustering algorithm. In face detection, there are two methods, we proposed, individually for application of single face and multiple faces detection in an image. About method of single face detection, we develop a face detection approach using a dynamic threshold function to improve Moghaddam’s Bayesian approach in YCbCr color space. The dynamic threshold function is defined based on Log-Sigmoid function with the number of the skin-color pixels in each image. Three processes with different thresholds are consecutively filtered the bitmap obtained from skin-color analysis of an original image to get the optimal face region. The process with the first threshold is used to roughly detect the whole human’s face, effectively fill all missing areas inside of the face and remove chips outside of the face. Then the process with the second threshold is used to erosion the protruding regions to become a similar oval-shaped as well as face shape. Finally, the process with the third threshold is used to extract the optimal face region and remove non-human skin color in the background. After filtering process Moghaddam’s Bayesian approach is used for face detection in the optimal face region. The experimental results show that our proposed method is more quickly and accurately to detect human’s face than the Moghaddam’s Bayesian method.
About method of multiple faces detection, there are many valid algorithms for face detection, but the size of filtering mask for detecting face regions is still a difficult problem now. In this thesis, we propose algorithms to provide the more suitable size of candidate face regions in complex images, which include more skin-color areas as arms, wear clothes, and background near the faces or some of these faces overlapped, for convenient and precise face detection. In color images, based on skin-color classification, the features classified is the first algorithm with a similarity-based clustering method (SCM) by two main features, both position and color. Then a frame integration algorithm is used for regions union if they belong to the same face. Furthermore, a frame segmentation algorithm can be used to partition of faces in the same region to generate the optimum of face boundaries. After performing the three algorithms above, the candidate face regions will be found by rejecting most framed regions if their size is too small or the ratio of height to weight is over than 2.3. Finally, recognition of the face regions in a color image is performed by an appearance-based method using spectral histograms as representation and support vector machines (SVMs) as classifiers.
In face recognition, an efficient face recognition system with low computational complexity, nonsupervised self-learning, and high-speed recognition is proposed for affine face image. The low computational complexity can be obtained by 2D wavelet analysis. A location coefficient variance feature vector (LCVF) based on wavelet analysis will significantly reduce the size of covariance matrix in principle component analysis (PCA). The nonsupervised self-learning, is a crucial property in the proposed system, can be obtained by using SCM. Moreover, a two-steps recognition process is derived to achieve rapid recognition. In first recognition step, a Bayesian analysis is used to detect the maximum likelihood of each Gaussian density of model classes in a large database. In turn, the detail recognition will be achieved by a normalized least square error method (NLSE). In the experiments, only one front face image of each person is used as model for recognition. And the results show that the proposed system is robust and has good recognition accuracy.

並列關鍵字

similarity-based clustering method (SCM) ； Bayesian analysis ； Wavelet analysisn ； principle component analysis(PCA)

參考文獻

[5] S. K. Singh, D. S. Chauhan, M. Vatsa, R. Singh, “A Robust Skin Color Based Face Detection Algorithm,” Tamkang Journal, Science and Engineering, vol. 6, no. 4, pp. 227-234, 2003.

[1] M. J. Zhang and W. Gao, “An Adaptive Skin Color Detection Algorithm with Confusing Backgrounds Elimination,” IEEE International Conference, Image Processing, vol. 2, pp. 390-393, 2005.

[2] C. F. Juang, H. S. Perng, and S. K. Chen, “Skin Color Segmentation by Histogram-Based Neural Fuzzy Network,” IEEE International Joint Conference, Neural Networks, vol. 5, pp. 3058-3062, 2005.

[3] M. Harville, H. Baker, N. Bhatti, and S. Susstrunk, “Consistent Image-Based Measurement and Classification of Skin Color,” IEEE International Conference, Image Processing, vol. 2, pp. 374-377, 2005.

[4] D. Chai and A. Bouzerdoum, “A Bayesian Approach to Skin Color Classification in YCbCr Color Space,” IEEE Proceedings, TENCON 2000, vol. 2, pp. 421-424, 2000.