基于双层特征融合的生物识别

Biometric Identification Based on Two-layer Feature Fusion

多模态生物识别可以弥补单模态生物识别存在的缺陷,已成为目前生物识别研究的主流趋势.现有的多模态生物识别大都使用传统的机器学习方法,而以深度学习为代表的新一代人工智能方法在该领域的应用研究相对较少.因此,提出了一种端到端、可训练的卷积神经网络(Convolutional Neural Network,CNN)模型用于多模态生物特征识别,并从单模态和多模态两方面研究模型结构和融合方式对识别性能的影响.在单模态识别中,研究不同网络层数和卷积核对识别性能的影响,并利用单模态识别的结果确定多模态识别的网络结构;在多模态识别中,为研究不同阶段特征融合对识别性能的影响,设计了两种不同的CNN结构;基于3种不同的融合方法,探索单层特征融合和双层特征融合机制对识别性能的影响,并通过组合优化给出一种最优的深度模型结构.为了评估本文方法的性能,分别在AR、Yale、Extended YaleB、LFW、PolyU和CASIA V1.0等6个标准数据库上进行验证.试验结果表明,基于CNN的单模态识别方法优于传统机器学习方法,本文提出的方法能够胜任单模态或多模态生物识别任务.

In order to make up for the defects of single modal biometrics,multimodal biometrics has become the main trend in biometrics research. However,most of the existing researches on multimodal biometrics use traditional machine learning methods,the new-generation artificial intelligence methods by deep learning are relatively few in this field. Therefore,this paper proposes a kind of end-to-end trainable convolution neural networks( CNN) model for multimodal biometric identification research. Specifically,the model structure and the way of fusion effect on the recognition performance are studied from two aspects of single modal and multimodal recognition. In the single modal recognition,the influence of different layer numbers and convolution kernels on the recognition performance are discussed,and the network structure of multimodal recognition is determined by the results of single modal recognition. In multimodal recognition,two different CNN structures are designed in order to check the influence of feature fusion in different stages on recognition performance. Based on three different fusion methods,the influence of single layer feature fusion and two-layer fusion mechanism on the recognition performance is explored. An optimal depth model structure is given by combinatorial optimization. To evaluate the performance of the proposed methods,extensive experiments are carried on six standard databases:AR,Yale,Extended Yale B,LFW,Poly U,and CASIA V1. 0. Experimental results show that the method based on CNN is superior to the traditional machine learning methods. In addition,the experimental results show that the proposed method can be competent for the single modal or multimodal biometric identification task.