具有普适性的手背静脉识别关键算法研究

Compared with other biological feature recognition methods, dorsal hand vein recognition becomes a research hotspot because of its multi advantages, such as living body recognition, internal characteristics and non-contact acquisition. And in the project the following three aspects will be researched. Firstly, in order to acquire high quality vein image which is helpful for extracting good feature, an image quality assessment function with characteristics of recognition object will be proposed, and vein image can be acquired through self optimization according to assessment result; Secondly, in order to improve feature extraction algorithm robustness for potential interference information in dorsal hand, image interference information judgment method will be proposed, and the feature vector of interference region can be corrected adaptively, so the proposed feature extraction algorithm will be more universal; Finally, during feature matching process, a recognition algorithm based on sparse representation with two regular expressions will be proposed, which not only restrains the influence of corrected feature errors in interference region, but also improves the universality of recognition algorithm for multi image acquisition conditions. In addition, in the project the three research contents, which include image acquisition, feature extraction and feature recognition, are linked together, and vein recognition key problems integration research can be achieved. With the implementation of the project, the theoretical research of biological feature recognition will be promoted greatly, and the core algorithms and techniques with more universality can be provided for real recognition system.

相对于其它生物特征识别方式，手背静脉识别具有活体识别、内部特征、及非接触式采集等优势，已逐渐成为研究热点。本项目拟在3个方面开展研究：① 为采集更有利于获取优质特征的高质量静脉图像，建立具有识别目标特点的图像质量评价函数，并基于评价结果实现静脉图像的自寻优采集；② 为提高特征提取方法对于可能出现的手背干扰信息的鲁棒性，提出一种图像干扰信息判别机制，并对干扰区域特征进行自适应修正，使提出的特征提取方法更具普适性；③ 在特征匹配过程中，提出一种基于双正则项的稀疏表示识别方法，不仅有效抑制干扰区域修正特征带来的误差影响，而且还能够增强识别对于多种采集条件的普适性。此外，本项目将以上图像采集、特征提取、特征识别3方面研究相互关联起来，实现了静脉识别关键问题的一体化研究。本项目的实施，将对生物特征识别的理论研究起到重要的推动作用，为真实识别系统提供了更具普适性的核心算法与技术。

手背静脉识别因其同时具有特定光源、活体识别、内部特征、及非接触式采集等优势，已成为生物特征识别技术的研究热点。针对手背静脉识别技术自身的特点，本项目在实施过程中主要研究内容如下：① 建立了包含142人手背静脉图像的数据集，并将其网上共享，为同行研究提供了重要资源；② 针对手背静脉图像特点，将非负矩阵分解理论引入至本项目中，并对非负矩阵分解模型进行优化与改进，获得了较高的识别准确率；③ 将一些深度学习方法引入至本项目中，其中包括卷积神经网络模型，YOLO模型等，进一步提高了识别的准确性。此外，针对四种不同的手背静脉图像数据集，算法具有较好的鲁棒性。④ 基于python与tensorflow框架开发了一套较为完整的手背静脉识别系统，能够将其应用在学生的出勤考查上；⑤ 手背静脉图像识别属于单标签标注问题，本项目由该问题引申提出了一种图像多标签标注算法，并在公开数据集上取得了较好的图像标注效果；⑥ 从私密性角度考虑，在本项目研究过程中提出了若干图像加密算法，在一定程度上保证了图像识别的安全性。本项目的实施，对生物特征识别的理论研究与实际应用起到重要的推动作用，为更多的图像识别问题提供了更具普适性的核心算法与技术。