【作者】 林森；

【导师】 苑玮琦；

【作者基本信息】 沈阳工业大学 ， 测试计量技术及仪器， 2013， 博士

【摘要】 在信息交互日趋频繁的现代社会，个人身份的认证显得越来越重要，并在电子交易，公共安全，商业金融等部门得到了高度重视和广泛应用。相比传统的身份认证方法，生物特征识别具有更高的可靠性和更好的简便性，同时它也是模式识别、图像处理、人工智能等领域的前沿方向，具有良好而广阔的发展前景。掌纹识别是生物特征识别领域的一种新兴技术，与其他生物特征识别技术相比，掌纹识别有许多独特的优势。此外，掌纹图像可通过非接触成像方法获取，具有非侵犯性，大众接受程度高，因此掌纹识别一直都是生物特征识别领域研究的热点之一。传统接触式采集装置便携性差，成本高，容易导致设备磨损、疾病传播，这使得非接触采集逐步成为主流的掌纹图像获取方式。然而，由于个人摆放习惯及感觉上的差异，可能导致在非接触采集时，出现掌纹图像变形、模糊等非理想状况，这无疑会降低掌纹识别系统的性能。针对以上实际应用中出现的若干问题，本文以非理想条件下掌纹识别为研究内容，以稳定特征为主要思想和方法，以自建的掌纹图像库为实验基础，提出变形和模糊掌纹识别问题的解决方案。具体来说，本文的主要工作如下：(1)针对非接触采集的低约束性可能会导致手掌摆放方式不定，与摄像头距离不定，从而引起掌纹图像变形的问题，充分考虑人手本身是非刚体的特点，从预处理角度提出基于Demons非刚性配准算法的变形掌纹归一化方法，即进一步增强变形掌纹与标准掌纹的相似性，弥补了传统刚性方法校正效果不佳的缺陷。首先使用改进的Demons非刚性配准算法进行变形掌纹的校正，再使用一系列技术指标进行效果评价，结果表明：在任取的图像序列内，各指标均优于传统的刚性配准方法，这意味着变形掌纹与标准掌纹的相似性得到了有效提高，同时方法具备优越性，为后续的特征提取和识别创造了有利条件。(2)针对传统尺度不变特征变换（SIFT）对非接触采集产生的仿射变形掌纹识别效果不佳的问题，从特征提取角度出发，利用一种改进方案，即仿射尺度不变特征变换（ASIFT），在未进行变形校正的情况下直接进行掌纹识别。首先建立了变形掌纹的仿射模型，模拟了相机光轴的经度角和纬度角，在仿射空间内提取图像特征。然后通过基于实际环境所建立的变形掌纹库验证算法性能，与SIFT算法及目前典型的掌纹识别方法进行对比。结果表明，ASIFT方法具备良好的抗掌纹仿射变形性能，证明了该方法能够成功解决掌纹变形问题，鲁棒性和稳定性强，具备优越性。(3)针对非接触采集时离焦状态导致掌纹图像模糊，造成识别系统性能降低的实际问题，在建立模糊模型并分析模糊现象机理的基础上，提出利用稳定特征进行模糊掌纹识别的思想，并由此提出基于离散余弦变换（DCT）和主线分块能量（PLBE）的模糊掌纹识别方法。首先使用DCT变换在频域内提取低频系数作为稳定特征，再使用改进的局部灰度极小值法提取空域内的稳定特征即主线，并使用分块方法计算主线能量形成特征向量，然后将频域和空域内的稳定特征进行融合，最后利用向量之间的欧氏距离进行识别。在自建的模糊掌纹库上的测试结果表明，与融合之前及其他典型识别方法比较，DCT+PLBE方法的识别效果为最佳，显示出该方法在识别性能上具备有效性和优越性，为解决模糊掌纹的识别问题提供了一条可行途径。(4)针对非接触采集时的离焦状况容易导致掌纹图像出现模糊的问题，秉承稳定特征思路，提出拉普拉斯平滑变换（LST）和手部几何特征（HGF）融合的模糊掌纹识别解决方案。首先使用LST变换提取模糊掌纹低频系数作为稳定特征，再提取手部几何特征，即手指相对长度和宽度作为测量特征，然后将LST特征和几何特征进行融合，最后利用特征向量之间的欧氏距离进行匹配和分类。在自建的模糊掌纹图库上的测试结果表明，与融合之前及其他典型识别方法相比，LST+HGF方法的识别性能最优，显示出方法具备可行性，能够进一步提升模糊掌纹识别系统的性能。更多还原

【Abstract】 In the modern society where the information interaction is becoming increasinglyfrequent, the identity authentication seems more and more important. It has raised greatattention and is being used extensively in the sectors of e-commerce, public security andcommercial finance. Compared with traditional personal identification methods, biometricshas higher reliability and simplicity. It’s the front direction of pattern recognition, imageprocessing, artificial intelligence or other fields, and has good and broad prospects fordevelopment. Palmprint recognition is a kind of new technology in biometrics field,compared with other biometrics technology, palmprint recognition has many uniqueadvantages. Besides, palmprint can be obtained by non-contact imaging method, which isnon-invasive with high degree of public acceptance. Therefore, palmprint recognition hasalways been one of the researching hotspots in the biometrics field.Traditional contact acquisition device has poor portability with high cost, and is easyto cause the wear of the equipment and the spread of the disease. All these make thenon-contact acquisition has gradually become the mainstream method for palmprint imageobtaining. However, the differences of the personal placing habits and feelings may resultin the non-ideal status with deformed or blurred palmprint images during the non-contactacquisition, which will undoubtedly reduce the performance of the palmprint recognitionsystem. In view of the above problems in the practical application, this thesis takes thepalmprint recognition of non-ideal conditions as the research content, the stable featuremethod as the main idea, and the self-built palmprint image database as the experimentalbasis, proposes the solutions to the deformed and blurred palmprint recognition problems.Specifically, the main work of this thesis is as follows:(1) Low restriction of the non-contact palmprint collection may cause different palmplacing gestures and different distance between the palm and the camera, these may resultin palm image deformation. Considering the non-rigid characteristic of human hands, anormalization model based on Demons non-rigid registration algorithm is proposed fromthe angle of preprocessing to better enhance the similarity between the deformed image and the standard image, and compensates the shortages of the traditional rigid methodwhich is not very effective. First, the improved Demons algorithm is used to normalize thedeformed palmprint; next, some technical indicators are employed to evaluate the results.The experimental results demonstrate that the indicators are better than the traditional rigidregistration method in randomly selected image sequence. This suggests that the similaritybetween the deformed palmprint and standard palmprint has been effectively improved,and the proposed methods have created favorable conditions for subsequent featureextraction and recognition.(2) In order to improve the limited effect of Scale Invariant Feature Transform (SIFT)about affine transformation caused by no parallel of the palm plane and the sensor planeduring contactless palmprint acquisition, a better method of palmprint recognition whichbased on Affine Scale Invariant Feature Transform (ASIFT) is proposed from the angle offeature extraction without deformation correction. Firstly, a model of affine deformation ofpalm is given, then the latitude and longitude of camera axis are simulated, and imagefeatures in the affine space are extracted. Based on the practical application environment,the deformed pamlprint database is established for the performance tests. Compared withthe SIFT and other typical palmprint recognition methods, the experimental results showthat the ASIFT can achieve the best performance when it’s employed to resist the affinedeformation of palmprint. In conclusion, the proposed algorithm can successfully solve thedeformation problem of palmprint, and it’s more effective, with superiority, robustness andstability.(3) In view of the practical problem of blurred image caused by defocus status fornon-contact palmprint collection which may reduce the performance of the recognitionsystem, a novel solution called stable features (SF) theory is proposed based onestablishing the blurred model and analyzing the blur mechanism, and a blurred palmprintrecognition method based on DCT and block energy of principal lines (PLBE) is presentedfurther. As the stable features, the low frequency coefficients are extracted by discretecosine transform(DCT) in the frequency domain, and the principal lines are extracted bythe improved local gray minimum method in the spatial domain. Thereafter the blockmethod is used for calculating principal lines energy to form the feature vectors. Then thestable features in the frequency and spatial domain are fused. Finally, the Euclideandistance between vectors is used for classification and identification. The experimentsbased on the self-made blurred palmprint database show that the proposed algorithm (DCT+PLBE) can get the best performance compared with no fusion and other typicalidentification methods, and this means it is an effective and superior approach which cansolve the problem of blurred palmprint recognition.(4) The defocus status caused by non-contact collection for palmprint will lead toimage blur, and result in the poor recognition performance of the identification system. Inorder to solve this practical problem, a novel solution is proposed based on the stablefeatures (SF) theory. Firstly, the laplacian smoothing transform (LST) is employed toextract the low-frequency coefficients of the blurred palmprint as the stable features.Secondly, the hand geometric features, namely the relative lengths and widths of thefingers are also extracted as the measurement features. Then the LST features andgeometric features are fused to constitute the new vectors. Finally, the Euclidean distancebetween the vectors is used for matching and classification. The experiments based on theself-made blurred palmprint database show, compared with no fusion and other typicalidentification methods, the proposed algorithm (LST+HGF) can get the best performance.This demonstrates that the algorithm is an effective and superior approach which canimprove the performance of blurred palmprint recognition system.更多还原