【作者】 郑苑；

【导师】 张星明； 史建新；

【作者基本信息】 华南理工大学 ， 计算机技术， 2010， 硕士

【摘要】 人脸识别作为一种主要的身份识别手段,在公共安全、视觉监控、数字身份认证、电子商务、多媒体和数字娱乐等领域有着广泛的应用前景。经过近四十年的发展,人脸识别技术已经取得了长足的进步,基本实现特定环境下的准确识别,并陆续出现了一些优秀的商用识别软件。但是在环境不可控的情况下,由于光照、姿态、表情、遮挡等变化的影响,已有的人脸识别算法性能大大下降,其应用范围受到较大的限制。其中,以光照问题给人脸识别带来的影响尤为显著。相关研究表明,同一个人的人脸图像在光照条件不同和经过各种主流的人脸识别方法特征提取后引起的差异,往往要大于不同的人在相同光照条件下的人脸图像的差异。因此,如何有效地对人脸图像进行光照纠正,以达到光照无关的预处理效果,是人脸识别研究的一个重要课题。为此,本文做了以下工作：(1)分析基于Retinex理论的光照预处理算法,并对算法进行改进：要消除光照对人脸图像的影响,在对数域执行高通滤波之前,先利用小波变换对图像进行高频强化。(2)为了提高人脸预处理识别率,将基于Retinex理论的W-G算法、直方图均衡化、gamma灰度校正三种预处理算法结合,寻找一种最优的组合方式。(3)分别使用PCA、LDA识别算法对上述光照补偿算法的预处理结果进行识别,结合各种光照补偿算法的特点对识别的结果进行分析,人脸数据库使用Yale B人脸数据库。比较分析两种识别方法的识别时间、识别率。实验结果发现,文中设计的五种光照预处理算法,在光照不理想的情况下均能取得较好的识别率；在正常光照情况下,不会引起图像质量的恶化。其中均衡化与W-G算法的分步处理得到的预处理效果最佳。直方图均衡化对灰度的均衡作用,削弱滤波中引起的光晕效果,同时高斯滤波滤去图像的低频分量,即与光照有关的入射分量,消除光照变化引起的图像的差异。更多还原

【Abstract】 As one of the most successful branches of personal identification, face recognition has great potential applications in public safely, visual surveillance, digital identity certification, e-commerce, multimedia and digital entertainment and other fields. The face recognition technology has developed rapidly over the past 40 years, basically accurate identification of a specific environment, and has been found in some excellent commercial recognition software. While the face recognition algorithm performance degrades under uncontrolled environment such as variant illumination, head poses, facial expressions, occlusion on face and so on. Research indicates that, in face recognition, variations caused by illumination are more significant than the inherent differences between individuals. Therefore, how to effectively correct the illumination face images to achieve the pre-independent effect of light is an important topic in Face Recognition.In our study paper, the main work can be summarized as follows:(1) Studies the illumination preprocessing algorithm based on Retinex theory, improve the algorithm as:to remove the effect of illumination on human face images, enhancing the high-frequency of the image using wavelet transform before the implementation of the log-domain high-pass filter with the image.(2) In order to improve recognition rate of face image preprocess, combining three preprocessing algorithms with:improved light compensation algorithm based on Retinex theory, histogram equalization and gray-scale transformation of gamma, find an optimal combination.(3) Identify the pre-processing results of illumination compensation algorithm with PCA, LDA recognition algorithm, using the Yale B face database. Combine the characteristics of illumination compensation algorithm to analyze the results of the identification, Comparative of two identification methods with identification time, recognition rate.The results showed that the five light preprocessing algorithm design in the test, achieve better recognition rate when light environment is not ideal; in normal lighting circumstances, the preprocessing algorithm does not cause the deterioration of image quality. Which equalization and improved illumination compensation algorithm sub-step processing is best handled by the pretreatment. Histogram equalization, balanced on the role of gray scale, weakening the halo effect caused by homomorphic filtering, and homomorphic filtering filters the low frequency components of image, eliminate the image differences caused by illumination change.更多还原