随着生物识别技术的发展，掌纹掌脉识别技术作为一种方便有效的身份认证解决方案受到了国内外学者的关注。本文围绕非接触式识别方式面临的难点和嵌入式平台上快速计算的瓶颈，以专项设备采集的掌纹、掌脉图像为处理对象，开展了图像预处理算法、图像配准方法、特征提取方法和识别策略等方面的研究。 (1)掌纹掌脉图像预处理算法：针对非接触式识别方式带来的背景不确定的问题，提出了一种基于高斯混合模型的手掌图像分割算法。采用高斯混合模型对图像边缘信息进行概率密度估计，对初始前景和背景建模，采用变分逼近的方法对模型参数进行估计。该方法综合利用了图像灰度信息的空间和聚类特性，不依赖专项设备，更具有一般性。实验证明该方法对红外图像和可见光图像的切割成功率分别达到97.8％和81％，在Cortex A8处理器上处理两幅图像的平均时间为5400ms；面对快速识别的要求和嵌入式平台计算能力的瓶颈，本文基于专项设备的特点构造了一种多模态快速手掌图像分割算法，采用SUSAN算子和最大类间方差法相结合的方法直接计算红外图像手掌关键点，并映射到可见光图像中，切割成功率达到了98.8％，平均运行时间仅170ms。 (2)图像配准方法：为了解决非接触式识别方式中手掌姿态不一致而导致的识别率下降的问题，提出了一种基于概率模型的点集匹配手掌配准方法。将回归分析和聚类分析相结合，引入转移变量建立起模型点集与场景点集之间的关系，对异方差噪声和离群点建模。在点集匹配不确定性估计时采用一种由粗到精的匹配策略。经过配准后类内误拒率明显下降，并且决策阈值越小时误拒率下降越大。 (3)掌纹掌脉特征提取：相对基于卷积计算的图像滤波方法，提高嵌入式平台上特征提取算法的执行效率尤为迫切。围绕竞争编码特征提取方法进行时间和效率的优化。利用迭代Gabor滤波器进行特征提取，在ARM9(三星2440)开发板上单幅图像特征提取时间由330ms左右降低为160ms左右。为了解决掌纹主线周围的纹线方向不准确和存在伪特征的情况，提出了一种基于二值加权竞争编码的特征提取方法，在校正纹线方向特征后结合有效值模板进行匹配。 (4)识别策略：设计了一种预匹配+重排序的快速匹配方案，使匹配效率提升了30％；设计了一个决策层融合的二维识别模型，将掌纹和掌脉匹配值融入一个坐标系下进行融合识别。 基于本文研究内容，搭建了一套基于嵌入式平台的非接触式掌纹掌脉识别系统，在Cortex A8处理器上基于1000人样本库采用1:N识别方式时一次识别成功平均时间700ms，完全可以满足实时性的需要。 关键词：掌纹识别；掌脉识别；融合；点集匹配；非接触式；嵌入式设备

With the development of biometric technology, as a convenient and effective authentication solution, palmprint and palmvein identification technology has attracted the attention of scholars at home and abroad. This paper focused on the difficulties of contactless recognition and the difficulties to acquire fast recognition speed based on the limited computing power of embedded platform. Based on the palm images captured using the device we developed, image preprocessing algorithm, image registration method, feature extraction method and recognition strategy were studied. (1)Palmprint&Palmvein preprocessing algorithm: Aiming at the problem of the background uncertainty brought by contactless recognition, a new method of palm image segmentation based on Gaussian mixture model was proposed. The Gaussian mixture model was used to estimate the probability density of image edge, the initial foreground and background are modeled using Gaussian mixture model, and the model parameters were estimated by the method of variational approximation. This method utilizes the spatial and clustering characteristics of image grayscale information, does not depend on our special device. The accurate segmentation rate is 97.8% for IR image and 81% for visible light image respectively, the average processing time on the Cortex A8 processor is 5400ms. In order to meet the requirements of fast recognition and the limited computing power of embedded device, a multimodal fast palm image preprocessing algorithm was proposed, which is based on the combination of Susan operator and Otsu's method to segment the edge of the palm in IR image first, and is mapped to visible light image subsequently, and the accurate rate reaches 98.8%, with an average processing time of 170ms. (2)Image registration method: In order to solve the problem of the palm gesture inconsistency in the contactless recognition, a robust image registration method was proposed using a point set matching method based on probability model. The two palms contour points sets to be registered was considered as model and scene respectively, the relationship between the model and the scene was established by a transfer variable, and the heteroscedastic noise and spurious outliers were modeled. The matching uncertainty was approximated by a coarse-to-fine variational inference algorithm. After registration, the FRR decreased obviously. (3)Palmprint&Palmvein feature extraction: Compared with the image filtering method based on convolution calculation, it is more urgent to improve the efficiency of the feature extraction algorithm on the embedded platform. The competitive coding method was optimized for both time usage and precision. The average processing time on ARM9 is reduced from 330ms to 160ms when the recursive Gabor filter is used to extract the feature of single image. In order to solve the problem of the palm lines direction uncertainty around principal lines and the problem of pseudo feature, a feature extraction method based on binary weighted competitive coding was proposed, which is used to match the direction corrected feature with the valid value template. (4)Recognition scheme: A fast matching scheme with pre-matching and reordering was designed, which improves the matching efficiency by 30%. A two-dimension recognition model was designed, which integrates palmprint and palmvein matching values into a coordinate system for recognition. Based on the research content of this thesis, a contactless palmprint and palmvein identification system based on embedded platform was built, and the average processing time reach 700ms on Cortex A8 processor based on the 1 :N recognition method, which can fully meet the real time needs. Key words: Palmprint Recognition; Palmvein Recognition; Fusion; Point Set Matching; Contactless; Embedded Device