【作者】 丛文龙；

【导师】 徐永红；

【作者基本信息】 燕山大学 ， 生物医学工程， 2010， 硕士

【摘要】 人脸识别是模式识别和人工智能领域的研究热点之一,在近些年受到了人们越来越多的重视。主成分分析算法(PCA)作为一种经典的人脸识别方法,识别过程简单直接,但识别效果受光照、表情等因素的影响较大;四元数主成分分析算法(QPCA)对图像像素之间的结构信息利用充分,但对图像的预处理要求较高,且大多用于彩色人脸图像识别。鉴于目前大部分人脸数据库收集的是灰度图像,本文以人脸灰度图像作为研究对象,综合考虑识别算法性能和计算复杂度两方面的因素,在特征提取和识别方面进行了对主成分分析算法的改进工作。首先,研究了小波包变换和四元代数的相关理论,提出了人脸特征的表示方法。利用小波包变换的正交多尺度特性,对人脸图像这一非平稳信号进行小波包变换,分解得到的各分量构造成四元数人脸模型。这样既保留了图像的全局信息,又获得了图像的细节信息,增强了各分量之间的相关性。然后,针对四元数主成分分析算法计算量大的问题,提出了二维四元数主成分分析方法(2D-QPCA),对四元数人脸模型进行降维并构造特征空间;在识别时,将特征空间划分为若干子块,对每个子块根据最近邻算法进行分类并对分类结果投票,根据投票结果实现最终的人脸识别。最后,在ORL、Yale、YaleB和Indian人脸数据库上进行实验仿真,研究了本文方法中参数与识别率之间的关系,并与PCA等算法进行了比较。实验表明本文方法改善了光照、表情变化的鲁棒性,具有识别精度高、计算复杂度低的特点。更多还原

【Abstract】 As one of hotspots in the field of pattern recognition and artificial intelligence, the research of face recognition has been drawn more and more attention in recent years. Principal component analysis (PCA) is a classical algorithm in face recognition with the character of that the process of recognition is simple and direct, but affected easily by illumination, facial expression and other factors; Quaternion principal component analysis (QPCA) is another face recognition algorithm, which can make full use of the structural information among pixels, but need high image preprocessing and often is referred to color face image recognition. Now a fact is that pictures collected in most people face databases are gray images. Considering two factors of recognition performance and computational complexity, this paper makes the improvements in feature extraction and recognition based on the PCA.Firstly, according to studying the theory of wavelet packet transform and quaternion, a new method of facial feature representation is proposed. Face images are decomposed by wavelet packet transform, which can provide multi-scale decomposition for a non-stationary signal. Each decomposition component is used to construct quaternion matrix, which can contain the global and detailed information of images, but also enhance the relationship between each component.Secondly, for quaternion principal component analysis has heavy work in calculation, two dimensional quaternion principal component analysis (2D-QPCA) is proposed to reduce dimensions and construct feature space. The space is divided into several sub-blocks, and each sub-block is classified based on nearest neighbor classification. The ultimate face recognition is completed according to classification results.Simulation experiments are carried on the face databases of ORL, Yale, YaleB and Indian to research the relationship of different parameters and recognition accuracy in the proposed method. Compared with PCA and other algorithms, experiment results show that the proposed face recognition method improves the robustness to illumination and expression changes, and is possessed the characteristics of high accuracy and low computational complexity.更多还原