超声波人脸识别方法研究
文献类型:学位论文
作者 | 苗振伟 |
学位类别 | 博士 |
答辩日期 | 2008-06-05 |
授予单位 | 中国科学院声学研究所 |
授予地点 | 声学研究所 |
关键词 | 超声波 人脸识别 宽带扫频信号 高分辨率距离像 功率谱 |
其他题名 | Ultrasonic Face Recognition |
学位专业 | 声学 |
中文摘要 | 使用超声进行人脸识别是一种不同于传统的可见光、红外线等方法的新兴的人脸识别技术,已经受到了越来越多的关注。相比于传统方式的人脸识别技术,超声人脸识别具有不受光线、温度影响的优势。本文即是在这样的背景条件下对超声波人脸识别技术展开研究的。 本文建立了人脸超声散射的简化模型,依据该简化模型分析了超声散射回波特性,并通过实验建立了大样本人脸表情数据库,经分类器识别获得了较高的识别正确率。 本文首先根据Freedman的水下高频声散射模型建立了简化的人脸超声散射中心模型,选取了具有较大带宽的线形扫频信号作为超声探测信号,可以实现较高的距离分辨率并能提高接收回波的信噪比,为超声波人脸识别的可行性提供了理论依据。在此基础之上,我们自主设计了一套完整的超声波人脸识别系统,采集了110人的多表情、多姿态超声人脸回波数据库,并分析了不同探测模式下噪声的影响,方便了我们后期的分类识别。后期的数据处理过程中,我们采用面部一维距离像和面部回波的功率谱作为超声波人脸识别的两类特征向量,并用多个分类器检测了它们的分类识别性能。当表情单一时,两类特征的识别正确率分别为99.4%和99.3%;当存在多个表情时,两类特征的识别正确率分别为98.8%和88.4%;当识别陌生表情时,利用一维距离像为特征,识别正确率最高为50.3%。通过多个通道融合的方式,可将识别正确率提升到90.5%。识别结果表明:在表情单一时,两类特征均具有较好的识别能力;当存在多个表情时,识别能力下降;数据融合的技术能够有效地提升超声波的人脸识别性能。 最后总结了本文的课题工作,指出了工作中的一些需要完善的地方,并对进一步提高识别正确率提出了一些想法。 |
英文摘要 | Recognizing faces with ultrasonic, which differs from the traditional ways used in face recognition such as visible light, infrared, etc. is a novel technology in the field of face recognition and has attracted more and more attention. Compared to the traditional ways mentioned above, its advantage is that it is not subject to light and temperature. In this paper, the feasibility of employing ultrasonic in face recognition is proved theoretically, a large database of face expressions is established by experiment and high recognition rate is achieved by classifiers. Based on the Freedman high-frequency scattering model, we build a simplified human face acoustic scattering model. In order to obtain a higher range resolution and improve the SNR of echoes, the chirp signal of large bandwidth is selected as the detection signal. With the self-designed face recognition system, we acquire a database consisting of 110 persons with varied expressions and gestures, and analyze the noise effect in different detection modes. High-Resolution Range Profiles (HRRP) of face and power spectrum of echo are extracted from the echo data as the two eigenvectors to test the recognition performance, the former one is used by Maximum Coherent Coefficient (MCC) classifier and Adaptive Gaussian Classifier (AGC), the latter one is used by Support Vector Machine (SVM) and AGC. The experiment result shows: in the case of single expression, the recognition performances using the two eigenvectors can reach maximum rates of 99.4% and 99.3%, respectively; in the case of multiple expressions, the recognition performances get maximum rate of 98.8% and 88.4%, respectively. As for recognizing unfamiliar faces, the recognition rate is only 50.3% maximally using one-dimension range profile as eigenvector by MCC classifier, so the method of multi-channel integration is employed to improve the performance, which can significantly raise the recognition rate to 90.5%. In the last part of this paper, we make a conclusion of our work, point out some areas that need improvement and put forward some ideas of obtaining a much higher recognition rate. |
语种 | 中文 |
公开日期 | 2011-05-07 |
页码 | 71 |
源URL | [http://159.226.59.140/handle/311008/368] ![]() |
专题 | 声学研究所_声学所博硕士学位论文_1981-2009博硕士学位论文 |
推荐引用方式 GB/T 7714 | 苗振伟. 超声波人脸识别方法研究[D]. 声学研究所. 中国科学院声学研究所. 2008. |
入库方式: OAI收割
来源:声学研究所
浏览0
下载0
收藏0
其他版本
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。