中文摘要：

指纹方向场对指纹的奇异点检测、特征提取和匹配、分类识别等至关重要。可靠地估计指纹方向场至今为止仍是一个具有挑战性的问题。现有方法一般先估计初始方向场，再对其进行去噪或者正则化处理。受最新的深度学习技术的启发，提出一种基于回归的端到端指纹方向场估计算法。该算法直接建立指纹图像块的纹理特征与其中心位置的脊线方向之间的映射关系。利用总变差模型分解指纹图像，以去除噪音的干扰；将指纹图像分成若干块，并利用深度卷积神经网络学习这些块的纹理特征与其中心位置脊线方向之间的回归函数。为评估文中算法的有效性，使用NISTSDl4数据库中的指纹作为训练数据，在FVC2002和FVC2004数据库上进行测试。实验结果表明：与已有的算法相比，该算法不仅简单易操作，而且具备较好的抗噪能力，可以准确地估计出奇异点及其周围的方向场，能够有效提高指纹识别率。

英文摘要：

Fingerprint orientation filed is crucial for fingerprint singularity detection, feature extraction and matching, classification and recognition, etc.. Many methods have been proposed for estimating fingerprint orientation field, mostly in two steps：estimation and regu- larization （or de-noising）. Yet, motivated by emerging deep learning techniques, a regression-based end-to-end fingerprint orientation field estimation method is proposed. It directly estimates the ridge orientation at the center of a fingerprint image patch through a regres- sion function from the texture feature of the image patch. Given a fingerprint image,the total variation model is applied to decompose it into cartoon and texture components. Then the texture component is divided into patches, using a Deep Convolutional Neural Network （DCNN） to estimate the ridge orientation at the center of each patch. The fingerprint images in NIST SD14 are adopted as training data to learn the DCNN-based regression function, and evaluate the proposed method on the FVC2002 and FVC2004 databases. The experi- mental results indicate that compared with the existing algorithms, the algorithm is simple and easy to operate, and has better anti -noise a- bility, which can accurately estimate the orientation field of singular point and its surroundings, and effectively raise the fingerprint recognition rate.