指纹是最常见和最热门的生物特征之一，被广泛应用于公共安全、刑事侦破、出入境管理等场景中。其中，指纹特征提取一直是研究热点问题。研究者们将常见的指纹特征按照不同尺度分为全局特征（如奇异点和方向场）、局部特征（如细节点）和细节特征（如汗孔点）三级。相比上述特征，指纹的姿态则很少被研究者们关注。在指纹识别中，姿态有着引入先验知识、减少特征匹配时搜索空间的用途。指纹的姿态规定了指纹内在的坐标系，提供了指纹的先验位置信息，从特征层级上可以看做指纹的第零级特征。本文将对指纹姿态提取问题进行研究，探求其与指纹全局特征的内在联系。本文的具体成果如下：1. 提出了一种针对现场指纹的姿态与方向场联合提取算法。针对现场指纹噪声较大、前景区域容易缺失等问题，本文在研究了指纹姿态和方向场的互补关系，提出了统一姿态下指纹方向场的两点假设，并采用大规模数据库搜索的方式进行了现场指纹的姿态及方向场联合估计。在现场指纹数据库NIST SD27 上，该算法的方向场和姿态估计结果均超过了当时的最佳结果。2. 提出了一种针对各类指纹图像的姿态与奇异点联合提取网络。指纹的姿态与奇异点息息相关，一方面，奇异点的位置决定了指纹的真实姿态；另一方面，在标准姿态下，奇异点也具有很强的空间分布规律。本文分别分析了现有的指纹姿态估计和奇异点提取算法，总结出其中的共通步骤，在此基础上设计了一种多任务学习的深度网络，实现了对各类指纹的姿态和奇异点联合估计。该算法在多种指纹数据库，包括NIST SD4、NIST SD14 滚动指纹数据库，FVC2004 DB1A 平面指纹数据库，NIST SD27 现场指纹数据库上都得到了更加准确的结果。3. 提出了一种基于平面指纹估计手指三维姿态的算法。手指三维姿态指手指按压在平面上的三维欧拉角度值。研究手指的三维姿态对于连接二维-三维指纹，以及丰富人机交互手段有着重要意义。本文采集了一批包含手指三维角度真值的平面指纹图像，依据指纹二维平面上的相对位置与手指三维角度的对应关系，基于深度期望回归思路设计并训练了一个深度神经网络完成由二维平面指纹对手指三维姿态的估计算法。在针对该任务而采集的指纹数据库上，该算法能够相对准确地完成手指三维欧拉角的估计。

Fingerprint is one of the most common and popular biometrics, which is widely used in public security, criminal investigation, entry and exit management and other scenarios. Fingerprint feature extraction is one of the hotspots in fingerprint recognition. Common fingerprint features can be classified into three levels according to their scales: the global level (such as singular points and orientation field), the local level (such as minutiae), and the very local level (such as sweat pores). Compared with these features, fingerprint pose has received little attention from researchers. In fingerprint recognition, pose has the advantage of introducing prior knowledge and reducing search space in fingerprint matching. The pose of a fingerprint is composed of the center and direction of the finger, stipulating the internal coordinate system of the fingerprint at the very global level, which provides the position information of the fingerprint. Thus, it can be regarded as the level-0feature of the fingerprint. In this thesis, the fingerprint pose estimation is studied and its internal relationship with level-1 features are explored. The specific contributions of this thesis are as follows:1. A joint extraction algorithm of pose and orientation field for latent fingerprint is proposed. Latent fingerprints are fingerprints detected at crime scenes. Due to overmuch noise and the absence of the foreground area of the fingerprint, many fingerprint recognition algorithms cannot process this kind of fingerprint well. To solve this problem, based on the dictionary-based orientation field estimation methods, this thesis studies the relationship between fingerprint pose and orientation field, discovers the global constraint of whole orientation field and proposes a joint estimationmethod for latent fingerprint pose and orientation field by means of exhaustive search on a large-scale database. The results of orientation field and pose estimation both exceed the state-of-the-arts on latent fingerprint database NIST SD27.2. A deep neural network for fingerprint pose and singular points is proposed. The pose of the fingerprint is closely related to the singular points. On the one hand, the definition of the fingerprint pose is determined by the position of the singular points. On the other hand, under the same coordinate system aligned by fingerprint pose, singular points in different fingerprints share similar space distributions. This thesis analyses the existing pose and singular points estimation algorithms, and summarizes the mutual steps. Based on the analysis, a multi-task deep neural network is proposed to estimate fingerprint’s pose and singular points at the same time. The relationship between these two features helps improve the performance. The proposed method performs better than the state-of-the-arts on different fingerprint databases, including the datasets of rolled fingerprints NIST SD4 and NIST SD14, the dataset of plain fingerprints FVC2004 DB1A, and the dataset of latent fingerprints NIST SD27.3. A deep neural network is proposed to estimate finger 3D pose via plain fingerprints. The 3D pose of a finger is defined by the Euler angles when the finger is pressed on a surface. The study of finger 3D pose is of great importance for connecting 2D and 3D fingerprints and human-computer interaction. In this thesis, the problem of finger 3D pose estimation via plain fingerprint is studied. To achieve this, this thesis collects a set of plain fingerprints along with ground truth 3D angles and trains a deep network with these data to complete the task of finger 3D pose estimation.The experimental results show that the proposed network can estimate the 3D pose of the finger precisely.