【作者】 亢静静；

【导师】 柯丽；

【作者基本信息】 沈阳工业大学 ， 信号与信息处理， 2011， 硕士

【摘要】 近几十年来,人脸识别技术已经取得了突飞猛进的发展,现今的研究人员主要集中在多姿态、多样本的人脸识别方法研究中。对于多姿态、多样本的人脸识别,每个人各种变化的样本获取是很困难的,然而每个人的单张正面照片是很容易获取的,因此研究单样本的人脸识别技术具有重要意义。本文提出了基于单样本注册的人脸视频图像识别研究的方法,其主要的工作如下：首先,对人脸识别技术的研究现状、发展历史、应用领域进行了论述,对常用的人脸识别方法做出了简要的总结。其次,本论文提出了利用Haar特征的人脸分类器进行人脸检测,在检测到的人脸上用双眼分类器检测出双眼,为了更加确切的分割出人眼区域,本文综合运用Gabor滤波器和眼睛模板(左眼和右眼的模板)在双眼区域内进行匹配的方法精确定位出眼睛(左眼与右眼)坐标,根据眼睛坐标在人脸上的几何分布对人脸图像进行几何归一化,然后再对几何归一化后的人脸图像进行灰度归一化,以便提高下一步的人脸识别的准确率。最后,采用基于小波特征融合的单样本注册人脸特征提取,进行分类识别。利用小波变换图像融合的方法,对注册的样本进行小波分解并把其低频信息存入库中,然后对测试人脸图像进行小波分解并提取其高频信息,把库中图像的低频信息与测试图像的高频信息进行融合,得到融合后的人脸图像,把融合图像与测试人脸图像之间的相容性(欧氏距离)作为分类特征,用SVM进行分类得出识别结果。用本文所提出的方法在部分FERET彩色人脸图库和自己用普通摄像头采集的十一组视频图像进行实验,达到良好的分类识别效果。大量的实验结果和统计分析表明本论文所提出的方法在光照、姿态、表情、饰物等影响下有很强的鲁棒性,并且具有较高的正确识别率和正确的拒绝率。更多还原

【Abstract】 Face recognition technology have improved by leaps and bounds in recent years. Now researchers focus on the study of multi-pose and multi-sample face recognition, but the method to obtain these images is very different. And the single front face image per person is easy to access. So it is very significant to study the face recognition with single training sample. This paper introduces the face recognition of video image based on single sample registered. The main works were as follows:First, it is discusses that the developing history, research situation and application fields on face image detection and recognition. This paper gives a survey of existing face recognition methods in recent years.Second, this paper introduces face detection using face classifier based on Haar features in the detecting images. When the face has been detected, using the trained binocular classifier extract the binocular region in face region. In order to more exact divide the eye areas, combine with Gabor filter and eye-template (left eye template and right eye template) matching accurately finds the eyes coordinates. According to the eyes geometric distribution in face, geometric normalize and grayscale normalize the face image to improve the face recognition rate.Last, recognize and classify the face images by single-sample registered feature extraction with wavelet transform. Using the methods of wavelet transformation and image fusion, obtain the low frequency information of registered image and deposit it to library and obtain the high frequency information of test image, then fuse the low frequency information in library and the high frequency to a fusion image. The compatibility (Euclidean distance) between fused image and test image as classification feature, classify the test image by SVM.To experiment on partial color face image in FERET and the 11 groups video image acquired by ordinary camera with the methods discussed above, achieved a good classify recognition. Extensive experiments and statistical analysis illustrates that these methods have strong robustness, high correct recognition rate and correct the rejection rate influencing by the illumination, posture, facial expression, decorations.更多还原