【作者】 魏丹；

【导师】 李树涛；

【作者基本信息】 湖南大学 ， 控制科学与工程， 2012， 博士

【摘要】 人脸识别是模式识别和计算机视觉领域的一个前沿课题,由于其具有非接触性、隐蔽性、易于理解以及图像采集设备成本低等优点,已经被越来越多的应用于安全监控、人机交互、人工智能以及电子商务安全中。本文以数字图像人脸识别技术为研究背景,在分析现有人脸识别方法的基础上,结合模式识别的最新理论,针对人脸识别中的表情、光照、遮盖等复杂情况,深入研究了基于稀疏表示和特征选择的人脸识别问题。本文的主要研究成果总结如下：1)基于图嵌入理论的特征选择方法。在图嵌入特征选择方法中,由于受到噪声影响,数据点的K邻近图稳定性会降低。针对这一问题,本文提出了基于特征分数的递归特征消除方法(FS-RFE)和基于子集水平分数的递归特征消除方法(SL-RFE)。在FS-RFE方法中,我们递归地移除具有最小特征分数的特征,并动态更新图的结构,以减少由于特征中存在大噪声而引起的负面影响。在SL-RFE方法中,通过迭代计算子集水平分数,递归地删除噪声特征,并更新图的结构。在UMIST及Yale人脸数据集上的实验结果表明,与θG-MFA,θG-LDA,θG-LSDF等特征分数方法相比,本文提出的FS-RFE和SL-RFE方法能够明显提高人脸识别的准确度,并显著提高算法对高维噪声的鲁棒性。2)基于链式采样的特征选择方法。针对非线性超高维问题降维,本文在特征生成机(FGM)方法基础上,提出了一种新的基于链式采样的特征选择方法。FGM方法在每次迭代过程中,特征根据分数进行排序,并且形成一个新的特征子集,当问题维数很高时,特征分数计算及其排序时间是无法接受的。而本文提出的方法通过特征采样方法加速计算,将稠密特征存入缓存器,并且舍弃稀疏特征,在迭代过程中将具有最大分数的一些特征保留在缓存器中,并逐步更新缓存器中的特征,形成链式采样,最后通过对缓存器的特征进行再排序,找到具有最大特征分数的一些特征做为有效特征,可以大大降低计算复杂度。在超高维数据集上的实验结果表明了本文提出的基于链式采样特征选择方法的有效性。3)基于工作集的快速有效稀疏表示求解算法。稀疏表示问题计算复杂度随着字典规模的增加迅速增加。为此,本文提出了一种求解稀疏表示问题的快速分解梯度投影算法(FDGP)。通过最小化一个有界约束二次规划问题来求解稀疏表示问题,在梯度投影迭代过程中,并不求解整个问题,而是选择梯度存在最大变化的元素作为工作集,从而将大规模优化问题转换为一些小规模有界约束二次规划问题来求解,既节省了内存消耗,又显著提高了大规模稀疏表示问题的求解效率,并最终提高了人脸识别的精度和效率。4)基于小波域稀疏表示的人脸识别方法。本文提出了基于小波域稀疏表示的人脸识别算法。由于小波高频子带可以捕捉小的细节信息而低频子带可以很好的表示轮廓信息,本文采用小波变换来分解人脸图像,建立包含高频和低频信息的多频字典,对高频子带和低频子带进行稀疏表示,通过计算高频、低频子带在多频字典下的拟合效果来进行分类。实验表明,即使当人脸图像存在着强烈光照表情变化或者小幅遮挡时,所提出的方法也可以对人脸进行有效准确地识别,从而提高了人脸识别的鲁棒性。5)基于决策融合的人脸识别方法。常用的特征级融合识别过程中信息存在相互干扰的可能性,容易造成融合结果性能的下降。为此,本文提出了基于决策融合的人脸识别算法。注意到局部二值模式可以反映图像的局部特征,线性判别分析可以充分提取图像的整体特征,本文首先对人脸图像进行多个尺度和方向的Gabor变换,再采用线性判别分析和局部二值模式两种方法提取Gabor图像的特征,采用K-最近邻方法进行识别,对得到的识别结果进行决策级融合得到最终结果。实验结果表明基于决策融合的识别方法结果准确率高于Gabor-LBP和Gabor-LDA方法,并且此方法随着实验测试人数(类别数)的增加,识别率保持稳定。6)基于分布式压缩传感理论的多传感器融合人脸识别方法。利用近红外和可见光图像的传感器内和传感器间的内在联系,提出了基于分布式压缩传感的多传感器融合人脸识别算法,将可见光与近红外人脸图像整体作为分布式压缩传感的多源测量信号,利用分布式压缩传感理论,将多源信号分解成共同部分和差异部分,采用共同部分对图像进行识别。共同部分有效地融合了近红外和可见光图像,既可以保持可见光图像容易采集,样本图像多的优势,又可以利用近红外图像对光照不敏感的特性,应用在人脸识别数据库上,取得了很好的效果。更多还原

【Abstract】 Face recognition is one of the cutting-edge research tompics in computer vision and pattern recognition areas. Because of its non-touchment and low-expense in system designing, face recognition has been widely applied in security surveillance, human-computer interaction, artificial intelligence, and electronic commerce etc. Targeting on the digital face image recognition, the negative effects that brought by the variety of light and expression, and the cover problems have been studied in the dissertation. By taking the more recent development in machine learning, this dissertation has thoroughly studied the face recognition based on sparse representation and feature selection. The main contributions of this dissertation can be summarized as follows:1) Graph embedding based feature selection.In the graph based feature selection, the stability of the K-nearest graph will degrade with the increasing noise features. In this dissertation, we developed two recursive feature elimination (RFE) methods using feature score (FS) and subset level (SL) score, respectively, for identifying the optimal feature subset. In FS-RFE method, we recursively remove the features with the least feature scores and update the graph with the selected features to reduce the negative influence on graph construction. In SL-RFE, we iteratively calculate the subset level score and recursively remove those feature with least scores based on the updated graph. The experimental results on UMIST and Yalefaces datasets verify that the proposed RFE method can achieve the state-of-art performance compared with the baseline methods such as θG-MFA, θG-LDA,θG-LSDF and SL, and can avoid the negative influence brought by the noise features on the graph effectively.2) Chain sampling methods for feature selection on ultrahigh dimensional problems.Regarding the dimension reduction in extremely high dimensional problems, in this dissertation, a sampling scheme is proposed to enhance the efficiency of recently developed Feature Generating Machines (FGM). In each iteration of FGM, the features are ordered by their scores to form a new feature subset. For high dimensional problems, the entire computational cost of feature ordering will become unbearable. Our method tries to keep those dense features in a buffer, drop those sparse features and speedup the algorithm using chain sampling on instances. In chain sampling, we just keep some features with the largest scores in the buffer when the iteration evolves and exchange the features in the buffer gradually. Finally, we reorder the features in the buffer and find the features with the biggest scores. Our proposed strategy can reduce this computational complexity significantly. Empirical studies on ultrahigh datasets on face image datasets showed the effectiveness of the proposed sampling method.3) Efficient large-scale sparse representation algorithm based on working set.The complexity of sparse representation will sharply increase with the scale of the dictionary. Regarding this issue, an efficient large-scale sparse representation algorithm, named fast decomposed gradient projection algorithm, is proposed in this dissertation for face recognition. In the proposed method, the sparse representation is addressed via solving a box-constrained quadratic programming problem. However, rather than solving the entire large scale problem, the proposed method selects those atoms with the largest absolute gradients as working set, which will transforms the original problem into a series of small box-constrained quadratic problems. By solving these small optimization problems, the large-scale sparse representation can be efficiently solved with very small memory requirements. The efficiency of the large-scale sparse representation can be greatly improved, which can make great improvements on the face recognition accuracy.4) Robust face recognition by sparse representation in wavelet domain.In this paper, we propose a novel robust face recognition algorithm by sparse representation in wavelet domain. Considering that the wavelet transform of an image can preserve its detailed and spatial distribution information it can be employed to extract the facial features. We construct multi-frequency dictionary which contains information of high frequency and low frequency, and obtain the sparse representation of high frencity and low frequency subband. Finally, we have the recognition result by compute the fitting of high frequency and low frequency subband in multi-frequency dictionary. The experimental results over two benchmark face databases demonstrate the robustness and improvements brought by the proposed algorithm.5) Face recognition method based on decision fusion.Most fusion methods on feature level needs to match different types of features. In addtion, because of the information collision of different types of features, the performance of the fusion results may be limited. To address this problem, a decision-level fusion method is proposed in this dissertation for face recognition. Notice that the local binary pattern (LBP) can reflect the local characteristics of the images, while the linear discriminant analysis (LDA) can efficiently extract the global image characteristics. Regarding this fact, we first do Gabor-transform on the images on multiple directions and scales, resulting in Gabor feature presentation of the face image. Then we extract the local features and global features by using the LBP and LDA, respectively. Finally, we fusion the recognition results from K-NN classification in decision level. Experimental results show that the proposed method based on the decision fusion shows superior performance than Gabor-LBP and Gabor-LDA method. More important, the proposed method shows stable performance over the increasing number of testing persons (testing classes).6) Face recognition method based on distributed compressive sensing of near infrared images and visible light images.By assuming that the infrared image and visible light image are sparse with respect to the whole image, we cast the near infrared image and visible light image of the same subject into an ensemble of inter-correlated image. To better capture the information of the two kinds of images to represent the near infrared and visible image of a given subject, we proposed to use the distributed compressive sensing to exploit the aforementioned sparsity of the assembled images. Finally, we proposed to do the image recognition based on the obtained distributed sparse coefficients, which is expected to obtain better performance than that with single near infrared image or visible light image. The experimental results on the benchmark dataset demonstrate the effectiveness of the proposed method.更多还原