手指静脉识别是一种利用手指掌侧浅静脉图像进行身份识别的技术，近年来得到广泛的关注和研究。由于手指静脉图像靠近红外光进行成像，成像时会受光照不均、皮肤组织散射、手指骨遮挡、手指的平移和旋转等因素的影响，采集的手指静脉图像质量比较差，所以，利用手指静脉图像进行身份识别，仍存在大量的问题需要解决。 一般来说，手指静脉识别包括图像采集、图像恢复、图像增强、感兴趣区域提取、特征提取、图像表示和匹配等技术。其中，特征提取和图像表示对手指静脉识别非常重要。基于手工设计的局部二值特征提取方法(如LBP、LLBP、PLLBP等)，具有简单、高效、对光照变化不敏感和对局部噪声鲁棒等优点，所以近年来得到了广泛研究，并成功应用在手指静脉识别任务上。但这类方法存在如下缺点：(1)设计特征需要领域知识。(2)特征捕获信息范围受限。因为提取一个局部特征采样的范围越大，该局部特征所捕获的信息就越多，但编码的维数就会太高，所需的存储和计算开销就会增大。(3)缺乏数据适应性，即特征提取过程不依靠整个训练数据。每个像素的局部二值特征提取是独立进行的，而不是靠训练数据进行学习得到的。(4)对成像时手指的旋转和平移敏感。针对手工设计的局部二值特征对成像时手指的旋转和平移不鲁棒，词袋模型被引入手指静脉识别领域，它能有效的组织从图像中抽取的局部二值特征，并在手指静脉识别任务上取得了较好的效果。但手工设计的二值特征的缺点也被继承下来。近年来，为了克服手工设计特征的缺点，基于学习的局部特征提取方法在其它图像识别领域中被提出，并取得了较好的识别效果。然而，现有的基于学习的局部二值特征提取方法忽视了训练图像的类别信息和训练数据的流形结构。另外，很少有算法期望学到的二值特征每类具有一定的结构。 本文主要针对上述问题，对基于局部二值特征的手指静脉识别技术开展了深入研究，主要的工作和贡献如下： (1)研究了一种基于定制的LLBP手指静脉识别方法。针对PLLBP特征各个方向上的分量对不同类的区分性不同，提出一种用PLLBP的不同方向分量对不同类进行身份识别的方法。通过训练，找出各个类最有区分性的几个PLLBP的方向分量。在测试时，用训练阶段找出的模板所属的类的最具区分性的方向分量计算测试图像的匹配得分。这样，不仅减少模板的存储开销，并且加快匹配速度，同时，提高了验证效果。 (2)提出一种基于监督学习局部二值特征的手指静脉识别方法。现有的基于学习的局部二值特征提取方法大都没有利用图像的类别信息，提取的特征区分性弱。针对这一问题，本文提出了基于学习的利用类别信息的手指静脉二值特征提取方法，并用于手指静脉识别任务上。 (3)提出了一种基于监督学习的个性化手指静脉二值特征提取方法。受LBP特征位一致性现象(同一类的LBP特征的某些位取值是相同的，即同时为0或者同时为1，而且，不同类的一致位的位置的组合是不相同的。)的启发，利用L₂,₁范数正则使属于每类的二值特征具有行稀疏特性，同时利用线性判别准则保证属于不同类的特征的行稀疏位置的组合不同。这样，学到的局部二值特征不仅区分性强，而且具有类结构(每类局部二值特征具有一定的结构，而且，不同类的结构不同)。这样，用所学的局部二值特征作为词袋模型的基特征，把手指静脉图像进行直方图表示，所获得的直方图表示也具有很强的区分性。本文在两个公开的手指静脉数据库上进行了实验，实验验证了所提出方法的有效性。 (4)提出了一种基于锚点非对称图流形学习二值特征的手指静脉识别算法。流形学习已经成功的应用在很多机器学习任务上，如降维、哈希和聚类等，而现有基于学习的局部二值特征提取算法很少利用原始数据的流形结构。这是因为局部特征学习所用的训练数据数量庞大，为这样庞大数据构建相似矩阵是一个运算和存储开销都比较大的操作。针对此问题，提出了构建锚点非对称图来近似相似矩阵的方法。这种方法在进行局部二值特征学习的过程中不仅有效的利用了训练数据的流形结构，还减少了存储和运算开销，在手指静脉识别任务上取得了比较好的效果。另外，考虑到基于监督学习的局部特征与基于非监督学习的局部特征之间存在差异，而且这两种特征可以进行互补，提出了一种基于图像表示级融合的手指静脉识别方法，并在两个公开的手指静脉数据库上取得了较好的识别性能。 关键词：手指静脉识别；二值局部特征；特征学习；流形学习；交替方向法。

 Finger vein recognition, which is a biometric technique using the vein pattern in finger palmer for human identification, has received the extensive attention and research in recent years. Because the finger vein image is imaged by the infrared light, there are uneven illumination, skin tissue scattering, finger bone occlusion, translation and rotation of finger when imaging, the quality of images in the existing databases are relatively poor. Therefore, there are still many problems that need to be solved in the field of the finger vein recognition. In general, the techniques of finger vein recognition include image acquisition, image restoration, image enhancement, region of interest extraction, feature extraction, image representation, and matching. Among them, feature extraction and image representation are very important for finger vein image recognition. Recently, handcrafted local features (e.g. LBP, LLBP and PLLBP), which are simple, effective, efficient, insensitive to lighting changes and robust to local noise, have been widely studied and successfully applied to finger vein recognition tasks. However, these methods have the following disadvantages: (1) Designing handcrafted local features requires domain knowledge. (2) The range of captured information is limited. Because the larger sampling range used for extracting a local feature, the more information is captured by the local feature, but the dimension of the binary feature will be too long and the storage and computational cost will be increased. (3) Lack of data adaptability. That is to say, the local binary feature extraction for each pixel is performed independently, rather than leaning through training data. (4) Sensitive to the rotation and translation of the finger when acquiring finger vein images. In view of the fact that such binary features are not robust to the rotation and translation of the finger, the bag-of-words framework is introduced into the field of finger vein recognition, which can effectively organize the local features extracted from the images and achieve better recognition results. However, the shortcomings of the handcrafted features have also been inherited. More recently, to overcome the shortcomings of handcrafted local feature, many learning-based local features are proposed in the other field of image recognition and has achieved good recognition results. However, the existing learning-based local feature extraction methods igore the class information of the training image, the manifold structure of original data during the learning process. In addition, few methods pursue the learned binary features with a certain class structure. This thesis focuses on the aforementioned problems and study deeply on the technology of extracting local binary features from finger vein images , and the main work and contributions are as follows: (1)A customized LLBP method for finger vein recognition was studied. Due to the component at different direction of PLLBP has different discrimination for different class, a method for identifying different classes using different directional components is proposed. Through training, the most discriminative components for each class are found. In the testing, the matching score of tesing image is calculated using the most discriminative components, which belong to the class of the template. In this way, not only the storage space of the template is reduced, but also the matching speed is accelerated, and at the same time, the verification performance is improved. (2)A based on supervised learning local feature for finger vein recognition method is proposed. Most of the existing learning-based local binary feature extraction methods do not utilize the class information of image, and the extracted features are lack of discrimination. To overcome this problem, this paper proposes a finger vein local binary feature extraction method based on the surpervised learning. (4)A supervised learning-based personalized finger vein local binary feature extraction method is proposed. Inspired by the LBP feature with bit consistency phenomenon (that is, some bits of the LBP codes belonging to same class have the same value, and the positions combination of these bits belonging to different classes is not the same), the L norm regulation is used to make the learned binary features belonging to each class sparse in row, and linear discriminant analysis is utilized to ensure that the combinations of spared row locations belonging to different classes are different. In this way, the learned binary featues are not only highly discriminative, but also have class structure (the features belonging to a class has a certain structure, and the structures belonging to different class are different). The image level representation using this binary feature by bag-of-words mode is also highly discriminative. In this paper, experiments were performed on two published finger vein databases, and the effectiveness of the proposed method was verified. (3)A based on unsupervised manifold learning local binary feature for finger vein recognition method is proposed. The manifold structure is very important for many machine learning algorithms, such as dimensionality reduction, hashing and clustering. However, the existing learning based local binary feature extraction algorithm rarely uses the manifold structure of the original data. This is because the training data used for local feature learning is big data, the constructing similarity matrix for these big data requires a large storage and computational overhead. To address this problem, this paper proposes a method using an asymmetric graph to appoximited to similarity matrix. This method not only effectively utilizes the manifold of the training data, but also reduces the storage and computational cost, and achieves better results on two public finger vein databases. Considering the complementarity between the featues based on supervised learning and the features based on unsupervised learning, a fusion method at the image representation level is proposed. The fusion method achieves good recognition performance on two public finger vein databases. Keywords: Finger vein recognition; Binary local feature; Feature learning; Manifold learning; Alternating direction method.