基于仿生模式识别及模糊理论的彩色图像人脸检测

【作者】 戚丽华；

【导师】 王耀明； 王淮亭；

【作者基本信息】 上海师范大学 ， 通信与信息系统， 2008， 硕士

【摘要】 人脸检测成为近年的研究热点,这主要来源于两大动力:学术价值和商业价值。这一重要课题的突破性进展将给人脸识别、表情姿态的识别、视频监控等相关领域的研究带来很大的推动作用,从而促进计算机视觉、模式识别等计算机科学分支甚至整个计算机科学的发展。目前对人脸检测的研究已经相当深入,值得一提的是,在脸部特征的选取上,目前为止,国内外的大量研究都集于五官中的“眼、鼻、嘴”的特征,而对眉毛的研究甚少。事实上,人类的眉毛具有足够好的稳定性、抗干扰性和多样性,而不会像眼睛和嘴巴那样受表情的变化而产生大幅的形变,因而可以作为一种独立的、良好的生物特征用于人脸检测。本文重点研究了彩色图像由眉毛检测实现人脸检测的算法。眉毛检测是人脸检测、识别的关键,其位置的确定可大致估计出人脸的尺度及方向。本文的眉毛候选区域的验证建立在眉毛定位的基础上。眉毛的验证使用了仿生模式识别方法,在此基础上被拒识的样本进一步采用模糊模式识别进行判别。具体来说,本文的工作重点分为以下几个方面:第一,依据眉毛处在人脸之中的条件,即眉毛由肤色围绕,为了减少搜索范围,对原图像进行肤色分割。提出了一种最优阈值分割方法分割肤色相似度图。较之固定阈值法,该方法较好地保留了肤色区域。对分割后的肤色区域,采用形态学滤波和欧拉数等原理进一步排除非人脸的肤色区域。第二,提出用眉毛检测来实现人脸检测的思路并作了探索性的研究。通过一个眉毛在人脸识别中所能起到的作用的评估性的实验,证明了眉毛在人脸识别中的地位不亚于眼睛等面部特征,这为本文工作的开展奠定了实验基础。为减少仿生模式识别验证次数,本文的眉毛初步定位算法利用眉毛的亮度及灰度投影等信息大致确定眉毛的区域。对于误检的区域,根据眉毛的几何规则排除之,如眉毛连线与水平线的夹角、检测区域外接矩形的宽与双眉中心连线之比等。第三,直接用人脸的全部数据进行训练的话,由于空间维数较高,训练时间长,检测耗时。本文先采集眉毛图像,并进行归一化,再提取矩不变量构成特征向量来进行训练。实际检测时,对算法检测出的眉毛区域进行尺度归一化再计算20阶Legendre矩,组成特征向量进行判别。这样相对于用整个人脸区域可大大缩短检测时耗。第四,对眉毛训练样本进行仿生模式建模,对待识别眉毛候选区域进行仿生模式判别。对该轮判决后遭拒识的样本进一步采用模糊模式识别方法,构造适当的隶属函数,通过实验确定判决阈值,从而实现对待识别眉毛的最终判定。从实验结果中可以看出,本文所采用的方法是较有效的,具有一定的理论和实用价值。更多还原

【Abstract】 Recent years, research on face detection has been very hot. This is because of its both academic and commercial value. The broken advancement of this important subject will bring great push to the fields such as face recognition, expression and pose recognition, video surveillance systems and so forth. And thus promotes the development of computer vision, pattern recognition and other computer science branch and even the development of whole computer science.Currently research on face detection has gone through deeply. What is worth mention is that by so far, most relative research on facial feature extraction both home and abroad has focused on eyes, noses and mouth features and research on eyebrows is so less. In fact, human eyebrow has enough good stability, anti-interference and diversity. Their shape wont’t vary much like eyes and mouth so they can be a good independent biologic features to be used in face detection.This thesis focuses on the research of human face detection realized by eyebrows detection in color images. As is a critical step towards face detection and recognition, the positions of eyebrows are commonly used for estimating a human face’s scale and orientation. The results of eyebrows localization premise the verification of eyebrow region in this thesis. Biomimetic pattern recognition method is used in the validation of the eyebrows and face for its excellent performance of recognition. And the samples which are rejected in previous recognition step will be recognized again by fuzzy pattern recognition method. In particular, main contributions of this thesis detail are as follows:Firstly, based on the prior knowledge that two eyebrows are in the face, i.e. eyebrows are surrounded by skin, skin segmentation can be applied to reduce the search region. Here, a method of segmenting the skin likelihood grey scale image is proposed based on the optimum threshold technique. Compared with those using a fixed threshold value, it’s more effective. Morphology filter and euler number principle is further used on the previously segmented skin likelihood region to exclude non-face region.Secondly, an idea of realizing face detection through eyebrows detection is proposed and a research based on it is done. Through an experiment of evaluating the role that eyebrows play in human face detection which results in that eyebrows are more important than eyes to some extent, we benefit from it that our research has experimental foundation and it’s meaningful. To reduce the amount of biomimetic pattern verification, the algorithm utilizes luminance and gray scale projection information of eyebrows to further determine candidate regions. For those error-detected regions, geometry information of eyebrows, such as the angle of eyebrow pair and the ratio between the distance of two eyebrows and the width of the outer rectangle of detected region, can be the rules of removal.Thirdly, if we utilize the whole face image data to train, the detection procedure will be very time-consuming because of its high space dimension. So in this thesis we first collect eyebrows images, normalize them by scalar and then extract moment invariants to form the eigenvector for training. In the actual test procedure, the eyebrows region detected by previous algorithm is normalized by scalar and then 20 orders’Legendre moments are calculated, which forms the eigenvector for testing. By doing so, the detection procedure will take far less time than using the whole face data to form the eigenvector.Fourthly, we use biomimetic pattern recognition method to confirm the candidate eyebrow region based on modeling of the training samples. Those samples which are rejected to be recognized in this turn will be further determined by fuzzy pattern recognition. That means a proper degree function of membership in a category to be constructed and then to determine a threshold through an amount of experiment. In this way the final confirmation of the eyebrows will be give out.Experimental results show that the method of this thesis is effective. It’s useful both theoretically and practically.更多还原