跨模态掌纹特征提取研究

In the field of public security, authentication and supervision of identity information is increasingly important. Biometric technology is safer, more secure and more convenient than traditional identification methods. According to law enforcement agencies in the United States, 30% of the crime scene marks the crime, from the palm of the hand, rather than fingerprints. Therefore, applying palmprint recognition on mobile terminal will help the public security system quickly troubleshoot suspects, significantly improve the efficiency of handling cases, thus has an important application and social value. But up to now, there are no reports or systems, mainly because of the large differences between the contact and non-contact imaging mode, the large change of the palm posture, the slow speed of recognition and so on. Utilizing deep learning and image processing tools, this project will study characteristics of palmprint image and carry out the research work from three aspects: how to build a cross-modality model using deep learning, so a non-contact palmprint could be non-linearly converted to a contact palmprint image; how to extract palm key points with deep learning model, so hand gesture could be reliable estimated and an palmprint image could be aligned quickly; how to use deep learning model to extract binary features or implement image classification, thus palmprint matching speed could be accelerated. This project will combine theory with practice. A palmprint identification prototype system will be developed according to needs of the mobile terminal.

在公共安全领域，身份信息的认证和监管日益重要。生物识别技术比传统的身份鉴定方法更具安全、保密和方便性。据美国的执法机构统计，30%的犯罪现场留下的犯罪印记，来自于手掌，而非指纹。因此，移动执法终端加入掌纹识别将可以帮助公安系统快速排查嫌疑分子，大幅提高办案效率，有着重要的应用和价值。但目前为止，还未见相关的报道或系统问世，主要因为接触与非接触式成像模态差异大、手掌姿态变化大、识别速度慢等问题。本项目深入研究掌纹图像特点，依靠深度学习、图像处理等工具，从三个方面开展研究工作：研究如何利用深度学习跨模态建模，建立非接触式掌纹到接触式掌纹图像的非线性映射；研究如何利用深度学习模型提取手掌关键点，可靠估计手掌姿态并将掌纹图像快速对齐；研究如何利用深度学习模型建立掌纹图像高效的二值表示和图像分类，加快掌纹比对速度。本项目还将理论结合实际，针对移动执法终端需求，进行掌纹识别原型系统的开发。

在公共安全领域，身份信息的认证和监管日益重要。生物识别技术比传统的身份鉴定方法更具安全、保密和方便性。据美国的执法机构统计，30%的犯罪现场留下的犯罪印记，来自于手掌，而非指纹。因此，移动执法终端加入掌纹识别将可以帮助公安系统快速排查嫌疑分子，大幅提高办案效率，有着重要的应用和价值。本项目深入研究掌纹图像特点，依靠深度学习等方法，围绕掌纹特征提取与模态转换，开展如下研究工作：利用深度学习跨模态建模，建立非接触式掌纹到接触式掌纹图像的非线性映射；学习不同光谱下的掌纹图像变换关系，建立不同光谱下的掌纹变换模型；利用深度学习模型建立掌纹图像高效的二值表示，加快掌纹比对速度；并针对手机掌纹图像去噪问题，提出了用户可参与的可调节去噪神经网络，优化掌纹图像；最终，开发了一套基于手机的掌纹识别原型系统。.研究成果包括发表论文21篇，其中在TIP、TIM、TIFS、IJCV等国际SCI期刊上发表论文16篇，在ACM MM、ICCV等国际重要会议上发表论文5篇，授权发明专利3项，获得国家技术发明二等奖、电子学会自然科学一等奖等奖励2项，参与组织CCBR、PRCV等会议，培养博士后1名，博士生2名和硕士生4名。