【作者】 张超；

【导师】 李昆仑； 刘明；

【作者基本信息】 河北大学 ， 通信与信息系统， 2010， 硕士

【摘要】 数据分析是模式识别、机器学习等领域的理论基础。聚类是一种重要的数据分析方法,它是将物理或抽象数据的集合分组成为由彼此相似的对象组成的多个簇(Cluster)的过程。为了使聚类的结果更准确,需要利用数据本身的属性信息,并且采用一些方法来降低错误率。近几年逐渐兴起的半监督学习和核方法能有效解决这两个问题。人脸检测(Face Detection)是模式识别领域的一个典型应用问题,是指给定任意一幅图像,判断在这幅图像中是否有人脸出现,如果存在则在输入图像中确定所有人脸的位置、大小和姿态的过程。基于肤色特征和聚类方法的人脸检测研究目前引起学者的广泛关注。本文将半监督学习策略应用到聚类算法中来指导聚类过程,提出了一种半监督核聚类算法——SKK-均值算法。算法利用一定数量的标记样本构成Seed集,作为监督信息来初始化K-均值算法的聚类中心,引导聚类过程并约束数据划分;同时还采用了核方法把输入数据映射到高维特征空间,并用核函数来实现样本之间的距离计算。在UCI数据集上进行了数值实验,并与K-均值算法和核-K-均值算法进行了比较。基于SKK-均值算法,本文提出了一种肤色模型(Skin Color Model)算法。算法首先对输入图片进行预处理,然后将少量像素点进行标记作为Seed集,再使用SKK-均值算法对肤色像素进行聚类生成肤色簇。依照各个肤色簇的像素点的概率分布,可以得到肤色模型;同时,对肤色簇进行形态学处理,可以得到肤色区域,这些肤色区域可以作为候选人脸(Face candidate)输入到AdaBoost算法中,进行实际的人脸检测。实验表明,使用本文的肤色模型算法可以快速而又准确地得到肤色簇,进而得到肤色模型和候选人脸区域,从而实现人脸检测。更多还原

【Abstract】 Data Analysis is a significant problem in the Pattern Recognition and Machine Learning fields. A important meyhod of Data Analysis is the Clustering. It is the organization of a collection of patterns into clusters based on similarity. In order to reduce the error rate, the Semi-Supervised Learning (SSL) strategy and Kernel Method have been introduced into the clustering process.Face Detection (FD) is the process which searches faces in the input image, and gives its position, size and pose if there’s any. It is a typical problem in PR fields. Nowadays, The study of FD algorithm based on skin color feature and clustering method attract more and more attention of researchers.This dissertation explore a semi-supervised clustering algorithm called Seed Kernel K-means (SKK-means) which is inspired by the kernel method and Seeding strategy based on the classical K-means algorithm. The algorithm uses a certain ratio of data points as the Seeds to generate initial cluster centers, and maps the data into feature space using kernel method. Our algorithm, which can be easily implemented, compares with respect to the other algorithm such as K-means and Kernel K-means, on 3 UCI databases (IRIS, Crabs and New-Thyroid) in some numeric experiment.This dissertation also raise a Skin Color Model (SCM) algorithm of FD based on SKK-means algorithm. The algorithm pretreated the input image firstly, and labels some pixels as the Seed points, then clustered each pixels of skin color as the Skin Color Clusters (SCC) using SKK-means algorithm. According to the probability of the pixels in SCC, we can obtained the SCM, and we can also obtain the skin regions after morphologic processing of the SCC. These skin regions can be used as the face candidate to put into the AdaBoost algorithm to detect human faces in practicality circumstance. Our experiments shows that it is more rapidly and accurately of our SCM algorithm to obtain the SCC. Thus, the Face Detection can be finally implemented after we got the SCM and face candidate.更多还原