【作者】 赵艳茹；

【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】 人脸识别是一个典型的图像理解与模式分类问题,在信息安全、刑事侦破和出入口控制等领域有着广泛的应用前景。经过近几十年的发展,人脸识别技术取得了长足的进步。目前最好的人脸识别系统在受控条件下已经能够取得较好的识别性能,但测试和实践经验表明:非理想条件下(光照、表情和姿势变化等)的人脸识别技术尚未成熟。Gabor小波变换具有良好的空间局部性和方向选择性,是一种良好的人脸描述方法,所以被广泛应用于人脸识别领域。Gabor小波变换处理人脸图像有很多优势,但是计算Gabor特征时具有较高的计算复杂度并需要大量的存储空间。本文研究主要贡献在于:将四元数小波幅值相位特征应用到人脸识别,与传统Gabor小波变换比较,具有明显较低的时间复杂度,更有利于实际应用。本文首先研究了四元数小波变换的结构和紧标架、实信号的四元数小波变换及其相位特性,在此基础上提出了QWT幅值/相位特征人脸表示方法。该方法将四元数代数和小波理论结合,有良好的时移不变性,可以提取一个幅值和三个相位信息,从而对人脸表情及光照变化有较好的不敏感性。采用分块投票策略将幅值/相位矩阵分成若干子块,对每个子块根据最近邻原则进行分类,并对各子块分类结果进行投票以实现人脸图像的最终识别。与PCA比较,本文方法在Indian、Yale、ORL、YaleB扩展数据库和CMU-PIE数据库的识别率分别高出14.8%、31.5%、9.9%、13%和3.4%。在ORL数据库,本文方法识别率略高于LDA方法;在Indian、Yale、YaleB扩展数据库和CMU-PIE数据库,本文方法都具有更高的识别率。在FERET’96数据库四个测试子集(Fb、Fc、DupI和DupII),本文方法的识别结果分别为94.1%、56.7%、58.1%和47.1%。虽然EBGM在测试集Fc上识别精度高于本文方法,但是本文方法在其它三个子集上高于EBGM而且有较少的可调参数和较低的计算复杂度。选取FERET’96数据库中任意一幅人脸图像进行识别,像素大小为64×64时QWT完成特征提取的运行时间只有0.047s,而Gabor小波变换为0.566s。另外,采用七种不同的光照补偿方法在光照明显变化的YaleB扩展数据库和FERET’96数据库上对人脸图像进行预处理,验证了本文方法结合可控滤波器和基于离散余弦变换的归一化技术后的人脸识别率明显提高。研究结果表明本文提出的人脸识别方法不但可以获得很好的光照、表情以及姿势变化的鲁棒性而且识别精度比传统方法有了较大的提高,计算复杂度较低。该算法在图像配准、人脸检测和人脸特征定位中有潜在的应用价值。更多还原

【Abstract】 Face recognition is a typical image understanding and pattern classification problem, which has a great deal of potential applications in information security, criminal detection and access control. After the development of recent decades, face recognition technology has made considerable progress. At present the best face recognition system under controlled conditions have been able to achieve better recognition performance, however, evaluation results and practical experience have shown that: the face recognition technology is not yet mature under non-ideal conditions (illumination, expression and pose changes, et al.). Gabor wavelet transform has good spatial locality and orientation selectivity, which is a good face description method, so it is widely used in face recognition domain. Even though Gabor wavelet based face image representation is optimal in many respects, it is computationally very complex and memory requirements for storing Gabor features are very high. The main contribution of this paper is: comparing with the Gabor wavelet of traditional method, the quaternion wavelet magnitude phase features are used for face recognition and they have a significantly lower time complexity and are more conducive to practical application.Firstly, the construction and tight frame of quaternion wavelet transforms (QWT), real signal of quaternion wavelet transform and quaternion wavelet transform phase characteristics are studied, and then the quaternion wavelet transform magnitude phase representation is proposed. Combining the quaternion algebra and wavelet theory, this method is a near shift-invariant tight frame representation, whose coefficients support a magnitude and three phase, therefore it has robustness to expression and illumination changes. These quaternion magnitude and phases are combined and divided into several sub-blocks, and then each sub-block is classified by a nearest neighbor classifier. These sub-block classification results are voted to complete the ultimate face recognition.Comparing with the PCA method, the recognition rate of this method is respectively higher on the Indian, Yale, ORL, YaleB Extended face database and the CMU-PIE face database by 14.8%, 31.5%, 9.9%, 13% and 3.4%. The recognition rate of the proposed method is slightly higher than the LDA method on the ORL face database and outperforms the LDA method on the Indian, Yale, YaleB Extended face database and the CMU-PIE face database. The recognition rates of the four FERET’96 probe sets achieve 94.1%, 56.7%, 58.1% and 47.1%, respectively. Although EBGM has better recognition rate on the probe set Fc, our method is superior on other three probe sets. Moreover, the method has few free parameters to adjust and has much lower computation complexity. Any one of the FERET’96 face database images is selected, when the pixel of face image is 64×64, the running time of QWT feature extraction is only 0.047s, and Gabor wavelet transform reaches 0.566s. In addition, seven different illumination compensation methods are used for face image preprocessing on the YaleB Extended face database and FERET’96 face database. Combining the steerable filter and discrete cosine transform of normalization techniques, the recognition rate of this algorithm has improved obviously.The study shows that not only the proposed face recognition method can get a good robustness to the illumination, expression and gesture, but also the recognition performance has greatly improved comparing with traditional methods. This algorithm has potentials for various applications in image registration, face detection and facial feature location.更多还原