【作者】 郭秀梅；

【导师】 周卫东；

【作者基本信息】 山东大学 ， 信号与信息处理， 2014， 博士

【摘要】 生物特征识别技术是一种根据人的行为特征或生理特征进行身份认证的技术。在公共安全、电子商务、金融等领域有着非常重要应用。适合身份鉴别的生物特征需要满足普遍性、稳定性、唯一性和可采集性。现阶段,常用进行身份识别的生物特征主要有：人脸、指纹、掌纹、静脉、虹膜、步态、语音、签名等。在这些生物特征中,由于手掌的面积较大,含有丰富的主线,乳突纹、皱褶等信息,且不易仿制与磨损,具备“人人拥有,人各不同,长期不变”的特点,使得掌纹识别成为生物特征识别的新兴技术。掌纹识别技术现在还处于发展阶段,各方面研究有待于进一步的深化和完善,因此对掌纹识别技术进行深入研究是非常迫切和必要的。在此背景下,本文对掌纹识别算法进行相关的研究,围绕掌纹特征的提取、分类等展开了以下研究工作：(1)提出了基于水平扩张毯子维的掌纹识别算法,将毯子维引入到掌纹识别领域。本研究将分形维数引入掌纹识别中,分析对比了常见的分数维,并根据掌纹纹理信息的方向,对毯子维进行了水平、垂直方向扩张,发现水平扩张毯子维更适合提取掌纹的纹理特征。另外,毯子维具有多分辨率特性。随着覆盖层数增加,上表面和下表面与图像逐渐分离,分辨率相应的也会降低。多分辨率毯子维针对不同的表面计算不同的毯子维,然后组成毯子维向量。该算法只需要对掌纹有效区域进行灰度均衡化处理,即可提取掌纹毯子维特征。分类器采用的是多次循移位的归一化相关器,可以有效的避免手掌旋转或移动带来的影响,具有较好的鲁棒性。如果只使用一个毯子维时,即基于单层水平扩张毯子维的掌纹识别算法,在香港理工大学掌纹数据库(版本2)和中科院掌纹数据库进行了试验,等错误率分别为0.1%和0.5%。如果使用两个或两个以上毯子维时,即基于多层毯子维的掌纹识别算法,使用毯子维个数越多,等错误率越低。如果使用两层毯子维时,在香港理工大学掌纹数据库(版本2)上等错误率降低至0.04%,并且对训练样本的多少具有较强的鲁棒性。(2)本文根据掌纹识别系统的实际应用,提出了掌纹旋转和遮挡等低质量待测数据库并进行相应的算法研究。提出了基于SRC、RSRC、CRC和RCRC的掌纹识别算法。基于分类的稀疏表示有两部分构成：残差和系数的稀疏性。为了使问题变为凸优化问题,残差和系数的稀疏性度量一般采用1范数或者2范数。此时稀疏模型可以分解为四种：标准的稀疏表示(SRC)、鲁邦的稀疏表示(RSRC)、协作表示(CRC)和鲁邦的协作表示(RCRC)。本文详细的分析该四种模型以及它们的优化方法,并将该四种算法分别应用到掌纹识别中。SRC与CRC相比,虽然CRC采用了稀疏性较差的2范数来度量系数的稀疏性,识别率却优于SRC,并且识别速度大大提高。这表明在稀疏表示中,类间协作表示的作用大于系数的稀疏性,CRC具有较好的分类特性。SRC和CRC在残差度量时均采用了2范数,两者只能适用待测图像比较理想的情况。若待测手掌受伤或采集的手掌放置位置变化或者提取ROI时存在一定角度的偏转时,残差度量采用1范数比2范数更适合这样的低质量待测样本,RSRC和RCRC在低质量待测样本数据库上取得了较好的识别效果,尤其RCRC在遮挡面积高达50%时,仍可以取得近70%的识别率。(3)提出了基于HM_LBP和CRC的掌纹识别算法。局部纹理特征在特征提取中起着举足轻重的地位。局部纹理特征具有以下优点：首先,图像的大量统计信息蕴含于局部纹理信息之中；其次,相对全局图像,在一个局部尺度上光照变化带了的灰度变化可以近似看作均匀和连续变化的,受光照亮度变化相对较小；再次,局部纹理一般通过局部算子实现,而局部算子的研究方向是挖掘判别力强同时鲁棒性强的算子,可以很好地克服图像旋转带来的影响。局部二值模式(LBP)是常见的局部纹理特征提取方法,在人脸识别等方向取得了较好的效果,但是传统的LBP模式大量的选择敏感模式被舍弃,并且舍弃的非规则模式所含的信息随着半径的增大而增大。分层多分辨率LBP可以解决传统LBP的缺陷,对于它带来维数过高的致命缺点,本文采用PCA对该特征进行降维处理。鉴于协作表示分类在掌纹识别中的有效性,因此本文将HM_LBP和CRC相结合,提出了一种新的掌纹识别算法,该算法在香港理工大学掌纹数据库版本2取得了非常理想的识别效果。同时对于难于确定的PCA维数问题,本文通过大量实验得出当PCA维数是训练样本总数的0.15倍时,识别率可以达到最优值,解决了之前PCA维数需要多次验证的缺点。更多还原

【Abstract】 Biometrics is a kind of identification technology by using the person’s physiological characteristics or behavioral characteristics.It plays a very important role in many fields, such as public safety, finance, electronic commerce, etc. The biological characteristics which can be used for identification must satisfy such features as uniqueness, universality, stability and scalability. At present the main features for identification include face, iris, fingerprint, palmprint, gait, signature, voice, vein and so on. As a new nember of biometric family, palmprints have attracted considerable attention from many research teams due to rich, stable and unique features, such as principal lines, wrinkles, ridges, minutiae pointd, singular points and texture. Nowadays, palmprint recognition technology is still in the stage of development, and further research remains to be deepened and perfect, so it is very urgent and necessary to make further study. Under this background, this paper made a study on palmprint recognition algorithm, and the works were summarized as follows:(1)In this paper we proposed a palmprint recognition algorithm based on horizontally expanded blanket dimension and it is the first time to introduce blanket dimension to palmprint recognition. We apply fractal dimension into palmprint recognition, analyze the common fractal dimension, and expand the blanket dimension both horizontally and vertically direction. The result showed that horizontally blanket dimension was more suitable for extracting palmprint texture features. In addition, blanket dimension has multi-resolution characteristics. As the coverage increases, the upper surface and the lower surface will separate from the image, correspondingly, and the resolution will become lower. Multi-scale blanket dimension is the blanket dimension vector which was composed by different blanket dimension computed according to different surface. The algorithm only needs gray equalization processing, and then can extract the blanket dimension features. To overcome the effect of rotation, the normalized correlation is calculated using three cycle shifts which has a good robustness. If we only use one blanket dimension, namely single-scale blanket dimension, and test the algorithm on Hong Kong Polytechnic University (PolyU) database (v2) and CASIA database. The equal error rate (EER) was0.1%and0.5%respectively. If we use two or more than two blanket dimension, namely multi-scale blanket dimension, the EER can reduce to0.04%on Hong Kong Polytechnic University (PolyU) database (v2). This suggests that if multi-scale HEBDs are employed, we can obtain a smaller EER, although the computational burden is increased slightly, and the EER was quite stable, irrespective of the size of training data.(2) In this paper we first construct low quality test database such as rotation or corrosion, and fill the gap of non-ideal palmprint image recognition. We proposed palmprint recognition algorithm based on SRC、RSRC、CRC and RCRC. Sparse representation consists of two parts:residual and sparsity of coefficient In order to make the problem into a convex optimization problem,1-norm or2-norm was usually adopted in terms of measurement of residual and sparsity of coefficient. The sparse model can be decomposed into the following four models:sparse representation based classification (SRC), robust sparse representation based classification (RSRC), collaborative representation based classification (CRC) and robust collaborative representation based classification (RCRC). This paper made a detailed analysis of these four types of models and their optimization methods, and applied them to palmprint recognition respectively. Compared with SRC, although the CRC adopted the poor sparse2-norm to measure residual, but the recognition rate of CRC is more preferable and its recognition speed is greatly improved, which shows that collaborations between classes work better than coefficient sparsity. However, because2-norm was used in the residual measurement are only used on the ideal condition that there is no much difference between the train and test images. If the image to be measured covers a certain area (an injured palm) or rotates with a certain angle (Deflection acquisition palm position change or ROI was extracted from certain angle), residua has better robustness when using1-norm than2-norm. So RSRC and RCRC have a stronger robustness on the low-quality image. Especially the RCRC algorithm, when the block area is as high as50%, it also can reach a70%recognition rate.(3) Proposed a palmprint recognition algorithm based on HM-LBP and CRC. Local texture feature plays a vital role in the feature extraction, which has the following advantages. Firstly, a large number of statistical information of images was contained in the local texture information. Secondly, gray change at the local scale changes with the light can be approximated as a homogeneous and continuous change, and the influence of illumination brightness is relatively small, compared with global image. Thirdly, local texture was generally obtained through the local operator, while local operators aim at discovering the ones of strong discrimination ability and robustness, which can overcome the influence of image rotation.Local binary pattern is a common method of extracting the local texture feature, which performs well in face recognition. But the traditional LBP pattern results in the loss of some information. In addition, the percentage of information loss increased with the increase of the radius value. HM-LBP can retrieve useful information from non-uniform patterns, but it brings about the high dimensions of palmprint features. This study reduced the dimension of hierarchical multiscale LBP features with PCA. CRC can achieve high accuracy of face recognition (FR) with significantly low complexity. HM-LBP and CRC are applied together in palmprint recognition in this study. The study is conducted with the Hong Kong Polytechnic University (PolyU) database (v2) in order to test the feasibility and performance of the algorithm. The results indicate that the proposed algorithm is simple and effective with high speed and100%accuracy of recognition. We also conclude that the best PCA dimension is0.15times the number of the total training samples with extensive experiments.更多还原