节点文献
基于环形对称Gabor变换的人脸识别算法研究
Face Recognition Approaches Based on Circularly Symmetrical Gabor Transforms
【作者】 王娜;
【导师】 王汇源;
【作者基本信息】 山东大学 , 信号与信息处理, 2015, 硕士
【摘要】 人脸识别技术作为计算机视觉和模式识别领域的研究热点,是一个具有重大理论研究意义和巨大应用价值的研究课题,涉及到人工智能、图像处理、神经网络等多方面的内容。随着社会的发展与科技的进步,社会对人脸识别技术的需求日益增加,准确而快速的人脸识别技术有着越来越广阔的应用前景。人脸识别是一个复杂而具有挑战性的研究课题,复杂性来自两个方面,一方面是人脸自身的变化,如人脸姿态和表情的变化以及年龄不同引起的变化,另一方面是背景的变化,如光照的变化。特征提取作为人脸识别中的关键技术,是人脸识别技术的研究核心,也是研究难点。好的特征提取技术要满足以下两点:①增强算法对姿态、表情、光照和尺度等变化的鲁棒性;②考虑到算法整体性能,计算复杂度不能过高。本文研究的重点也是放在如何快速提取有效特征上。因为环形对称Gabor变换相对于Gabor变换具有数据冗余度小和严格的旋转不变性的优点,所以本文选用环形对称Gabor变换来提取特征。论文主要工作为:对CSGT进行了深入的研究,提出了多种基于CSGT的多尺度特征融合方案,在以上研究的基础上提出了基于CSGT的人脸识别方法,基本思路就是对人脸图像进行CSGT,并在变换域进行多尺度特征融合,然后用子空间方法对融合后的特征进一步提取特征,最后用分类器分类。主要内容包括:(1)深入研究了CSGT,分析了不同参数对环形对称Gabor函数的影响,提出了5种基于CSGT的多尺度特征融合方案,并对融合后的特征进行分析,最后选择了3种识别效果较好的融合特征进行实验。(2)提出了基于CSGT+PCA+SVM的人脸识别算法,该方法先对图像进行CSGT,并用多尺度特征融合方法进行特征融合,选用经典的PCA方法对融合后的特征进一步压缩,最后用SVM进行分类。在此方法中选用的多尺度特征融合方法为平均图法和最大值图法,用这两种方法融合后得到的特征图像与原图像大小一致。在ORL人脸库和FERET人脸库上进行实验,实验结果证明了本文所提出的基于CSGT的多尺度融合特征的有效性及本文所提出的算法的可行性。(3)对一种模板大小随尺度变化的CSGT进行了研究,并将其与固定模板的CSGT都用在了实际实验中。(4)提出了基于CSGT+2DPCA+NN的人脸识别算法及基于CSGT+2DPCA+SVM的人脸识别算法。在这两种算法中,我们先用多尺度特征融合方案构造特征图像,然后用2DPCA方法进一步提取分类特征。在CSGT+2DPCA+NN的人脸识别算法中,我们对可变模板的CSGT和固定模板的CSGT分别做了实验。在基于可变模板的实验中根据融合方案的不同提出三种方法即*CSGT1+2DPCA、*CSGT2+2DPCA、*CSGT3+2DPCA,它们对应的融合方法分别为最大值图法、平均图法和5尺度上下连接的融合法。在基于固定模板的实验中,我们只对使用最大值图法融合得到的特征图进行了实验,该方法记为CSGT1+2DPCA。在ORL人脸库和FERET人脸库上进行实验,并将4种方法与己存在的其它相关方法进行了比较,通过比较得知,本文所提出的算法不但计算简单,而且能取得更好的识别效果。在CSGT+2DPCA+NN算法的基础上,提出了CSGT+2DPCA+SVM算法,并与CSGT+PCA+SVM方法进行了对比。在ORL和FERET人脸库上的实验结果表明,相比与CSGT+PCA+SVM方法,CSGT+2DPCA+SVM方法不但计算简单,耗时较少,而且具有更高的正确识别率。
【Abstract】 As a research hotspot of computer vision and pattern recognition, face recognition technology is a research topic which has important theory significance and application value, which involves the content of artificial intelligence, image processing and neural network etc. With the development of the society and the progress of science and technology, the demand for face recognition technology in society increases constantly, and the application prospects of accurate and fast face recognition technology become more and more broad. Face recognition is a complex and challenging research subject, where the complexity comes from two aspects:.on one hand the change of the face itself, such as the change of facial gestures and expressions, and changes caused by different ages, on the other hand, the change of the background, such as the change of illumination.The key part of face recognition is feature extraction. Good feature extraction technology must meet the following two points:① strong robustness of face recognition approaches to the variable factors which include illumination, expression, poses and the rotation;② low computational complexity. This thesis mainly studies face feature extraction. Compared with the traditional Gabor wavelet transform, the circularly symmetric Gabor transform (CSGT) has the advantages of low redundancy and strict rotation invariance, so CSGT is selected as the method of feature extraction in this thesis.The main works we have done include what follows. Sophisticated study is conducted to CSGT, and several multi-scale feature fusion schemes are proposed. Study is conducted to the template of CSGT. New face recognition methods are proposed based on CSGT. First, face images are mapped onto the CSGT domain and the feature images are constructed by using multi-scale feature fusion scheme in the same domain, then the features are further extracted by using subspace method, finally, classifier is used to achieve classification.The main work of this thesis is as follows:(1) In-depth research is conducted on CSGT:different influences of different parameters on CSGT are analyzed; 5 multi-scale feature fusion schemes are proposed and three fusion schemes which have good recognition effect are chosen to conduct experiment.(2) A new face recognition method based on CSGT and PCA+SVM is proposed. The face images are mapped onto the CSGT domain and the feature images are constructed by using multi-scale feature fusion scheme first. Then the features are further extracted by using PCA method. Finally, SVM is applied to achieve classification. The fusion schemes used in this algorithm are average figure method and maximum figure method respectively. After fusion, the size of the feature image is the same as that of the original image. Experiments on ORL database and FERET database are carried out. The experimental results show the effectiveness of the multi-scale fusion features and the feasibility of the proposed method.(3) Study on the variable template CSGT is conducted and experiments based on fixed template CSGT and variable template CSGT are performed respectively.(4) Two new approaches based on CSGT and 2DPCA are proposed, which are CSGT+2DPCA+NN and CSGT+2DPCA+SVM respectively. In this part,2DPCA method is used to extract classification features. In the CSGT+2DPCA+NN approache, experiments based on fixed template CSGT and variable template CSGT are performed respectively.CSGT+2DPCA+NN face recognition method:According to different fusion scheme, three methods based on variable template CSGT are put forward, which are *CSGT1+2DPCA, *CSGT2+2DPCA and *CSGT3+2DPCA, correspond to 3 different fusion schemes:maximum figure method, average figure method and images of 5 scales mosaic algorithm. Since the maximum figure method has the best recognition results, the method CSGT1+2DPCA based on fixed template CSGT are proposed too. Experiments on ORL database and FERET database are carried out.The experimental results show that the four methods proposed in this thesis can all achieve better recognition results than existing approaches.CSGT+2DPCA+SVM face recognition method:In this algorithm, the Support Vector Machine (SVM) is used as the classifier. Comparative experiments with CSGT+PCA+SVM indicate that the CSGT+2DPCA+SVM method can save calculation time and have higher recognition rates.