婴幼儿指纹识别若干关键问题研究

Recognition of infants and young children is of great importance for their healthy growth. Fingerprint is the most suitable biometric for recognition of infants and young children. But due to the problems of weak fingerprint signal, large growth deformation, small finger skin area, and low cooperation when capturing their fingerprints etc., the existing fingerprint recognition algorithms cannot be directly applied to infants and young children. How to improve the fingerprint recognition performance for infants and young children is still a challenging and important research topic. This project aims to develop new theories and methods based on the deep learning framework to address several key problems of fingerprint recognition for infants and young children. First, multi-layer deep neural networks will be constructed to learn the features and context information for fingerprint quality assessment and learn the transforms from the low to high quality fingerprint images for fingerprint reconstruction and enhancement. Second, a growth deformation model will be built based on recurrent neural networks to model the complex changes of fingerprints caused by the rapid growth of infants and young children. Finally, for fingerprint feature extraction and matching, the multi-scale convolutional neural networks will be studied for detection of fingerprint feature points. In adddition, the 2 channel multi-layer convolutional neural networks will be constructed to learn the multi-level features of fingerprints with joint optimization of feature extraction and matching. The research results are expected to improve the fingerprint recognition performance for infants and young children. The research of this project has academic innovation and important application value.

婴幼儿身份识别对其健康成长具有极其重要的意义。指纹是最适合婴幼儿身份识别的生物特征，但是婴幼儿指纹信号微弱、生长形变大、指端皮肤面积小、采集配合度差等问题，现有指纹识别算法无法直接应用于婴幼儿群体，如何提高婴幼儿指纹识别率仍然是一个挑战性的重要研究课题。本项目旨在基于深度学习框架，对婴幼儿指纹识别若干关键问题，开展新理论、新方法研究。首先针对指纹图像质量评估和增强，构建多层网络模型，自动提取指纹特征和上下文信息评估图像质量，学习从低质量到高质量指纹的变换关系重建指纹图像。然后构建基于递归神经网络的指纹生长形变模型，对指纹随婴幼儿快速成长而引起的复杂变化进行建模。最后，针对婴幼儿指纹特征提取和匹配，研究基于多尺度卷积神经网络的指纹特征点检测方法，构建双通道卷积神经网络，自动学习指纹多层次特征，联合优化特征提取和匹配。研究成果有望提高婴幼儿指纹识别性能。本项目研究具有学术上的创新和重要应用价值

婴幼儿身份识别对其安全和健康成长有着非常重要的意义。指纹是最适合婴幼儿身份识别的生物特征，经过众多研究者几十年的努力，指纹识别技术研究已取得很多重要成果，但是婴幼儿指纹信号微弱、生长形变大、指端皮肤面积小、指纹采集配合度差等问题，现有自动指纹识别技术无法直接应用到婴幼儿群体，如何提高婴幼儿指纹识别率仍然是一个挑战性的重要问题。近几年发展起来深度学习理论的深入研究和应用，为婴幼儿指纹识别提供新的思路和方法。本项目针对婴幼儿指纹识别的若干关键问题，基于深度神经网络框架，研究新的指纹图像处理和分析方法，并对这些方法的理论和应用进行了较充分的研究和分析。.首先，针对婴幼儿缺失公开指纹图像库问题，采用两种指纹采集器，开展0-6岁婴幼儿指纹图像数据采集，构建了一个含有6476张指纹图像数据库，为后续算法研究提供重要数据支撑。然后，针对婴幼儿指纹图像分辨率低问题，提出了一种基于密集连接金字塔的生成式对抗网络（DCP-GAN）模型，重建高分辨率指纹图像，提高婴幼儿指纹识别准确率，研究成果发表于国际生物特征会议IJCB2021会刊（EI检索）。其次，针对低质量指纹图像分割和增强问题，提出了基于低尺度嵌套U型卷积神经网络的分割算法、基于分块DenseUNet的增强方法，有效去除背景噪声、增强纹线清晰度，提高指纹图像质量和识别准确率，研究成果发表于国际知名期刊IEEE Trans. on Information Forensics and Security,2021,16(1):1709-1719等。最后，针对指纹特征提取和匹配，提出了基于多任务全卷积神经网络的特征点提取算法和基于图卷积神经网络的匹配算法，该方法不仅能准确检测特征点和计算特征点的方向，而且能有效融合指纹局部和全局特征信息，提高指纹识别准确率，研究成果发表于国际知名期刊Pattern Recognition,2021,120:0-108189等。.本项目研究的指纹图像处理和识别方法，在婴幼儿指纹图像库和其他低质量指纹图像库上进行验证，能有效提高指纹识别的鲁棒性和准确率。相关研究成果可以进一步开发成具有自主知识产权的指纹识别软件，为婴幼儿特殊群体提供一套快速、准确、廉价的身份识别解决方案，可用于跟踪婴幼儿免疫接种情况、打击拐卖儿童犯罪和户籍管理等国家重大需求领域。