基子稀疏表示的人脸识别算法研究

发布时间：2018-05-08 18:48

  本文选题：人脸识别 + 稀疏表示　； 参考：《山东大学》2016年博士论文

【摘要】：作为一种生物特征,人脸具有可随身携带、不会丢失、不易被盗取等优点,而且人脸图像采集方式友好、无需配合甚至具有隐蔽性。基于人脸的身份识别将成为未来身份认证和识别的主流方法,在经济、民用、军用、公安等领域具有广阔的应用前景,是当前基于生物特征身份识别领域中的研究热点。目前非限制条件下的人脸识别技术还不完善,还需进行深入研究并提出高效的识别算法。稀疏表示算法具有较高的分类性能,在图像分类及模式识别领域得到了广泛研究和应用。构造过完备冗余字典和稀疏表示矢量的快速求解是稀疏表示理论应用中的两大问题。论文针对人脸识别这一典型的小样本问题,围绕基于训练人脸图像的冗余字典构造方法、稀疏表示矢量的快速求解算法以及进一步提高稀疏表示分类性能等问题展开了以下研究工作。(1)提出利用位平面图像协作表示分类投票决策的人脸识别算法。研究了正则化最小二乘协作表示(CRC_RLS)分类算法的原理,分析了其与稀疏表示分类(SRC)的区别与联系以及它们在识别性能上的差异,在此基础上提出利用人脸图像的位平面信息和投票决策算法对协作表示算法进行改进,以提高协作表示分类算法的准确性和稀疏表示算法的识别速度。为突出人脸图像的轮廓,增强图像中的识别信息,本算法采用累积分布函数对图像进行直方图均衡。对均衡图像进行位平面分解,得到包含不同类别信息的多个二进制位平面图像,同一位平面图像构成同一位平面数据库。本文采用256级灰度图像进行实验,一幅人脸图像可分解出8幅位平面图像,一个灰度人脸图像库扩展为8个不同位平面人脸图像库。利用协作表示分类算法分别训练出每个位平面的正确识别率。实验证明图像均衡后,第2个位平面和第8个位平面具有相同的正确识别率,第3和第4位平面的正确识别率均在0.25以下。选择正确识别率较高的1、5、6、7、8等5个位平面,分别利用协作表示算法对输入图像进行识别。采用“最高票当选”制对5个识别结果进行投票。当多数投票结果不唯一时,定义为投票决策失败,进行二次判决：利用各位平面图像构造虚拟加权图像,在虚拟加权图像库上再次利用协作表示算法对图像进行分类。利用二进制位平面图像代替二值灰度位平面图像,避免了在虚拟加权人脸图像中低位平面被高位平面“淹没”的问题,权重系数由位平面的“位序数”和“正确识别率”共同确定,既保证了不同位平面具有不同的权重,又限制了高、低位平面的权重差值不会过大。在ORL和FERET人脸库上,正确识别率分别为97%和98%,与CRC_RLS算法相比,均提高了4%。与SRC算法相比,分别提高了4%和2%。训练图像的位平面分解、权重系数的训练与虚拟加权训练图像的构造均在训练阶段完成,不会影响实时人脸识别速度,因此识别速度与CRC_RLS相当,比SRC算法的识别速度提高了10倍以上。(2)提出基于局部构造模式近邻样本协作表示的人脸识别算法。随着字典原子数量增多,稀疏表示矢量的求解复杂度急剧上升,从而引起识别速度下降,同时字典原子与测试图像在结构上的相似度越高,正确识别率也越高。论文提出一种基于LCP特征的自适应字典学习算法,利用该字典对测试图像进行协作表示分类,既提高了识别速度,又提高了正确识别率。LCP特征包含局部结构(LBP)和局部微观构造(Mic)两层特征。本算法提出利用卡方系数的负对数定义基于归一化直方图卡方距离的“x2-LBP-相似性,,和基于bin-比例直方图距离的‘'χ2-BRD-LBP-相似性”,用以衡量图像在LBP特征层的相似程度,采用欧氏距离判断图像在微观构造特征层上的相似度(MiC-相似性),通过大量实验给出了χ2 -LBP-相似性和χ2-BRD-LBP-相似性的合法阈值、近邻阈值以及MiC-相似性的合法阈值、近邻阈值的经验取值范围。详细分析讨论了直方图特征与MiC特征级联／并联、bin-比例直方图特征与MiC特征级联／并联四种融合方法下错误拒绝识别、错误接收识别以及近邻样本选择三个问题。联合LBP特征和Mic特征,确定图像在LCP特征上的相似性,并根据预先设定的近邻阈值自适应地选择近邻样本,构造冗余字典。本算法提出的冗余字典的原子数量降到了训练图像数量的2/3以下,而且原子结构与测试样本的结构具有更高的相似性,正确识别率得到提高。基于比例直方图自适应构造的冗余字典对有遮挡图像的分类具有更高的鲁棒性。在ORL、FERET、YaleB和AR人脸库上,无遮挡识别时,本算法比SRC RLC的正确识别率提高了3%左右,利用AR库进行围巾和墨镜遮挡实验,正确识别率可达到85%。(3)提出利用Gabor近邻和压缩感知降维进行稀疏表示分类的人脸识别算法。人脸识别属于典型的小样本问题,利用人脸图像作为原子构造的字典不满足原子数量大于特征数量的条件,以上两种算法在使用协作表示分类之前,首先对图像进行PCA降维以满足字典要求,特征选择的问题依然存在。若采用压缩感知对人脸图像进行降维,则可避免特征选择难题,克服人脸识别中的小样本问题。本算法提取人脸图像的低维Gabor特征,在Gabor特征空间,利用相关系数自动寻找测试样本的近邻并作为表示基构成表示矩阵。通过大量实验证明了合法数据与训练样本的平均相关系数不依赖于具体的测试样本,并给出合法测试数据相关系数阈值的经验值。对合法测试数据,以“类平均相关系数”为准则选择近邻样本并构成表示矩阵,表示基涵盖了训练样本的全部类别,同时表示矩阵中表示基的数量减少了一半,而且表示基和测试图像具有更高的结构相似性,更符合压缩感知理论对于表示矩阵的要求。采用随机分布的高斯矩阵作为感知矩阵对人脸图像进行感知,将高维人脸图像投影到任意低维的观测空间上进行识别。分别采用正交匹配跟踪算法(OMP)和线性规划优化算法求稀疏表示矢量,并逐类完成测试样本的种属判决。与SRC算法相比,识别速度提高了5倍,在无遮挡的识别中,正确识别率提高了5%,对于AR库上的围巾遮挡和墨镜遮挡,正确识别率提高了将近1倍,分别达到83%和73%。
[Abstract]:As a biological feature, human faces have the advantages of portable, not lost and uneasy to be stolen, and face image acquisition is friendly, without coordination or even concealment. Face based identification will become the mainstream method of identification and identification in the future, and it is widely used in the fields of economy, civil, military, public security and so on. With the prospect, it is a hot topic in the field of biometric identification. At present, the face recognition technology under non restrictive conditions is not perfect. It needs to be studied and put forward the efficient recognition algorithm. The sparse representation algorithm has high classification performance. It has been widely studied and applied in the field of image classification and pattern recognition. The two major problems in sparse representation are constructed over complete redundant dictionaries and sparse representation vectors. This paper focuses on the typical small sample problem of face recognition. This paper focuses on the redundant dictionary construction method based on training face images, fast algorithm for sparse representation and further improvement of sparse representation. The following research work has been carried out on class performance. (1) a face recognition algorithm using bit plane images to represent classification voting decision is proposed. The principle of regularized least squares cooperative representation (CRC_RLS) classification algorithm is studied, and the difference and relation with the sparse representation classification (SRC) and their difference in recognition performance are analyzed. On this basis, the cooperative representation algorithm is improved by using the bit plane information and voting decision algorithm of the face image to improve the accuracy of the cooperative representation classification algorithm and the recognition speed of the sparse representation algorithm. The algorithm uses the cumulative distribution function to highlight the contour of the face image and enhance the recognition information in the image. Carry out histogram equalization. The image is decomposed by bit plane, and multiple binary bit plane images containing different types of information are obtained. The same bit plane is composed of the same plane database. In this paper, a 256 level gray image is used to experiment. One face image can decompose 8 bit plane images and a gray face image library is expanded. 8 different plane face images are developed. The correct recognition rate of each bit plane is trained by cooperative representation classification algorithm. The experiment shows that after the image is balanced, second bit planes and eighth bit planes have the same correct recognition rate, and the correct recognition rate of both third and fourth bit planes is below 0.25. 1,5,6,7,8 and other 5 bit planes, using the cooperative representation algorithm to identify the input images respectively. The 5 recognition results are voted by "the highest vote". When the majority of the voting results are not unique, it is defined as the vote decision failure, and the two decision is made: using the image to construct the virtual weighted image, in the virtual weighted image. The cooperative representation algorithm is used to classify the image again. The binary bit plane image is used instead of the two value gray level plane image to avoid the "submergence" of the low level plane in the virtual weighted face image. The weight coefficient is determined by the "number of bit numbers" and "correct recognition rate" of the bit plane. The weight difference between the different level planes is different and the weight difference of the low level plane is not too large. The correct recognition rate is 97% and 98% on the ORL and FERET face database respectively. Compared with the CRC_RLS algorithm, both the 4%. and the SRC algorithm are improved, and the bit plane decomposition of the 4% and 2%. training images is improved and the weight coefficients are trained and virtual. The construction of weighted training images is completed in the training stage, which does not affect the speed of real-time face recognition, so the recognition speed is equal to that of CRC_RLS, which is more than 10 times higher than that of the SRC algorithm. (2) a face recognition algorithm based on the cooperative representation of the nearest neighbor samples is proposed. With the number of dictionary atoms increasing, the sparse representation vector is used. The complexity of the solution increases sharply, which causes the decline of the recognition speed, and the higher the similarity between the dictionary and the test images, the higher the correct recognition rate. This paper proposes an adaptive dictionary learning algorithm based on the LCP feature, which uses the dictionary to classify the test images collaborating, not only improves the recognition speed, but also improves the recognition speed. The correct recognition rate.LCP features include the local structure (LBP) and the local micro structure (Mic) two layers. The algorithm uses the negative logarithm of the chi square coefficient to define the "x2-LBP- similarity" based on the normalized histogram of the square distance of the normalized histogram, and the '2-BRD-LBP- similarity' based on the bin- proportional histogram distance from the bin-, so as to measure the image in LB The similarity degree of P feature layer is measured by Euclidean distance (MiC- similarity) on the microscopic structural feature layer. Through a large number of experiments, the legal threshold of chi 2 -LBP- similarity and X 2-BRD-LBP- similarity, the nearest neighbor threshold and the legal threshold of MiC- similarity, the range of empirical value of near neighbor threshold are discussed in detail. Histogram features and MiC features cascade / parallel, bin- proportional histogram features and MiC features cascaded / parallel four fusion methods for error rejection recognition, error reception recognition and near neighbor samples selection three problems. Combine LBP features and Mic features to determine the similarity of the image on the LCP feature, and according to the predetermined nearest neighbor threshold self threshold. The number of redundant dictionaries proposed by this algorithm is reduced to less than 2/3 of the number of trained images, and the structure of the atomic structure is more similar to the structure of the test sample, and the correct recognition rate is improved. The redundant dictionary based on the proportional histogram adaptive construction has the occlusion image. The classification has higher robustness. On the ORL, FERET, YaleB and AR face library, this algorithm improves the correct recognition rate of the SRC RLC by about 3%, using the AR library to carry out the shawl and dark mirror occlusion experiment. The correct recognition rate can reach 85%. (3) and the sparse representation of face recognition using the Gabor near neighbor and the compressed sensing reduction is proposed. Face recognition is a typical small sample problem. The dictionary that uses face image as an atomic structure does not satisfy the condition that the number of atoms is larger than the number of features. The above two algorithms first reduce the dimension of the image to satisfy the dictionary before using the cooperative representation classification, and the problem of feature selection is still existing. If pressure is used. Contraction perception can reduce the dimension of face image, avoid the problem of feature selection and overcome the small sample problem in face recognition. This algorithm extracts the low dimension Gabor feature of face image. In the Gabor feature space, the nearest neighbor of the test sample is automatically searched by the correlation coefficient and the representation matrix is formed as the representation basis. A lot of experiments prove that the method is valid. The average correlation coefficient between the data and the training sample does not depend on the specific test sample, and gives the empirical value of the correlation coefficient threshold of the legal test data. The number of medium representation is reduced by half, and the representation base and the test image have higher structural similarity, which is more in line with the requirement of the representation matrix in the compressed sensing theory. The Gauss matrix of random distribution is used as the perceptual matrix to perceive the face image, and the high dimension face image is projected on the arbitrary low dimension observation space. Identification. The orthogonal matching tracking algorithm (OMP) and linear programming optimization algorithm are used to obtain the sparse representation vector, and the test sample is completed by class by class. Compared with the SRC algorithm, the recognition speed is increased by 5 times. The correct recognition rate is increased by 5% in the unshielded recognition. The correct recognition rate for the shawl occlusion and sunshade occlusion on the AR library is correct. Increased by nearly 1 times, reaching 83% and 73%., respectively

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41
，

本文编号：1862604