基于CNN的手指三模态多级特征融合与识别

CNN-Based Finger Trimodal Multi-Level Features Fusion and Recognition

随着人们对生物特征识别的要求不断提高,手指因为包含有丰富的生物特征信息受到越来越多的关注。手指同时携带了指纹、指节纹和指静脉三种模态信息,作为一个整体来表征个体的生物特性,具有得天独厚的优势。本文中,我们提出了一种基于CNN (Convolutional Neural Network)的手指三模态多级特征融合与识别算法,以融合指纹、指静脉和指节纹三种模态的特征信息来进行身份识别。该融合框架由联合优化的三部分组成,手指三模态特征提取模块,手指三模态浅层特征融合模块和手指三模态多级特征融合模块。手指三模态特征提取模块由三个并行的特征提取框架组成,分别用于提取手指单模态多级特征。浅层特征融合模块以CNN作为融合框架,以手指三模态特征提取网络的浅层输出特征作为输入,来获取三模态浅层融合特征。多级特征融合模块以全连接网络作为融合框架,手指三模态浅层融合特征和手指三模态的深层特征作为全连接网络的输入,以实现手指三模态多级特征融合。最后在手指三模态数据库上进行了大量实验,本文提出的算法识别精度可以达到99.91%。实验结果表明本文提出的手指三模态多级特征融合与识别算法性能优于其他现有的算法。

With the increasing demand of biometric recognition, finger, which contains abundant biometric information, has played an important role in the identification recognition field. There are finger-print (FP), finger-knuckle-print (FKP) and finger-vein (FV) on different parts of a finger. As a whole, finger has unique superiority to explore multimodal fusion recognition technology. In this paper, we propose a finger trimodal multi-level features fusion framework based on CNN, which integrates feature information of finger-print, fingervein and finger-knuckle-print for identification. The proposed trimodal multi-level features fusion framework consists of finger three-modal features extraction module, three-modal shallow features fusion module and three-modal multi-level features fusion module, which are jointly optimized. The finger three-modal features extraction module is composed of three parallel and independent CNN frameworks, which are used to extract multilevel features of FP, FV and FKP separately. The three-modal shallow features fusion module uses CNN as fusion framework, and takes the shallow features of the three biometric patterns as inputs to obtain shallow fusion feature. The multilevel features fusion module takes the fully connected network as the fusion framework. The trimodal biometric shallow fusion feature and the deep features serve as the inputs of the full connection network to realize trimodal multi-level features fusion. Experi-mental results show that the recognition accuracy of the proposed algorithm can be as high as 99.91%. A comprehensive analysis of many experimental results shows that the proposed trimodal multi-level features fusion framework performs better than other state-of-the-arts.