Fusion Recognition Method Of Multi-modal Biometrics Coding

In recent years,finger-based multi-modal fusion recognition has attracted much attention for personal identification.However,the recognition performance of multi-modal biometrics is usually sensitive to illumination and pose variation.Hence,exploring suitable method of feature representation is especially significance for developing finger multi-modal recognition system.In this thesis,we propose a local coding-based multi-modal fusion recognition method.1)A feature coding method based on symmetric cross-weighted local graph structure(SCW-LGS)is proposed for finger-vein recognition.Oriented Gabor filters are firstly used for venous region enhancement.Then,a multi-orientation coding algorithm is proposed to locally represent the position and gradient relationships among the pixels in a neighborhood of the Gabor enhanced images.2)A finger multi-modal fusion method based on feature coding is proposed.A bank of Gabor filters are first used for direction feature enhancement in finger images.Then,an improved coding scheme is developed for finger features representation.Finally,the coded tri-modal images of a finger are divided evenly into non-overlapping blocks,and the histograms of all blocks are concatenated into a feature vector for similarity matching.3)A feature extraction method combining local coding algorithm with convolutional neural network(LC-CNN)is proposed.Firstly,three traditional coding algorithms are reconstructed by a bank of convolution filters.Then,the pre-trained CNN modal is used to further extract the primary features of finger images.Finally,the extracted feature vector is input into the SVM to implement image classification.Based on this,two fusion strategies are proposed to achieve finger multi-modal fusion recognition.Experiments are implemented to verify the recognition performance of the proposed methods.The experimental results show that the proposed approaches achieve excellent recognition performance.It can effectively solve the problem of finger posture rotation,and improve the robustness of finger feature representation.