【作者】 徐佳杰；

【导师】 徐勇；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2010， 硕士

【摘要】 在近二十年中自动人脸识别引起了人们的广泛关注,人们迫切希望计算机能拥有和人一样的强大的依据人脸来识别人身份的能力从而使世界变的更加智能。但是,到目前为止,自动人脸识别还远未达到人们的期望;尤其是在复杂情况下,传统的基于人脸图像间相似性计算的自动人脸识别技术的性能难以达到应用要求。例如,脸部存在部分遮挡且遮挡比例稍大的情况下,自动人脸识别技术一般会出现识别错误或拒识。又如,研究表明,在光照条件存在较大变化的情况下,依据已有方法得出的同一人脸的不同图像间的差异性可超过不同人脸的图像间的差异性,此时必然会得出错误的人脸识别结果。如上因素也是目前人脸识别技术未能如指纹识别技术一样得到普及的原因。稀疏描述方法被引入到人脸识别领域后带来了人脸识别的突破性进展。已有的一些实验说明,在很多情况下包括图像存在噪声或被部分遮挡的情况下,稀疏描述方法可取得大大优于其他方法的分类正确率。实际上,基于稀疏描述方法的人脸识别技术是一个全新的方法学,它为人们解决已有的技术难点提供了一个新颖的视角和思路。但是,目前稀疏描述方法的图像识别应用仍有诸多问题需要解决。例如,现有的稀疏描述方法只利用了部分训练图像来表达测试图像,但是人们不知道究竟哪些训练图像被利用而哪些没被利用。又如,由于人脸图像的维数非常高,现有的稀疏描述方法的效率将非常低,在大容量训练集上尤其如此;该缺点使得方法不能应用于实际。本文针对如上问题,设计了几个新颖的基于稀疏描述思想的人脸识别方法,这些方法不仅计算效率高,而且能得到很高的识别精度。为解决现有的可见光人脸识别系统易受光照变化影响的问题,本文还设计并实现了一个基于人脸近红外与可见光图像的人脸识别系统。系统具有优异的性能。更多还原

【Abstract】 In the past two decades face recognition has attracted much attention. People hope that the computer can have a powerful capability of using the face to recognize the identity as we can do and the world can become more‘intelligent’. However, up to now, automatic face recognition does not reach this goal of people. Especially under the complex condition, the performance of conventional face recognition technique that is constructed on the basis of the computation of the similarity between the face images cannot meet the requirement of real-world applications. For example, when the face is partially occluded, conventional face recognition technique usually incorrectly recognizes or rejects to recognize the face. The research also shows that under the condition of greatly varying lighting condition, the difference of the images of the same face obtained using conventional face recognition methods might be greater than the difference of two images respectively generated from two different faces. Consequently, this will lead to incorrect recognition result. Indeed this factor is also one of the reasons why the automatic face recognition technique is not as popular as the fingerprint recognition technique.After the sparse representation method is applied to face recognition, a breakthrough has been made. Previous experiments show that under many cases including the case where the face is partially occluded or there is noise, the sparse method can obtain much higher classification accuracy than other methods. Indeed, sparse-representation-based face recognition technique is a novel methodology and it provides us a novel idea and viewpoint that is very useful for addressing the technical problems in face recognition. However, when the sparse representation method is applied to image recognition, there are still some unsolved problems. For example, previous sparse representation methods exploit only a portion of the whole training image set to represent the test image, but it is not known which training images are exploited and which are not. Moreover, because the dimension of the face image is very great, the computational efficiency of the sparse representation methods will be very low especially in the case where the training sample set has a large size. This is disadvantageous for the real-world applications. In order to solve the above issues, we devise several novel sparse-representation-idea-based face recognition methods. These methods not only are computationally efficient but also can obtain high accuracies. To address the issue that the performance of previous visible-light-based face recognition systems is affected by the varying illumination, we devise a face recognition system that combines visible light images and infrared red images of the face to robustly classify the face. This system has an excellent performance.更多还原