Research On Convolution Kernel Transformation And Matching Distances Fusion For Finger-Knuckle-Print Recognition

Biometric recognition(face,fingerprint,palmprint,etc.)has been a new and promising way for identity authentication.Compared with other modalities,fingerknuckle-print(FKP)contains lots of advantages,such as rich identification features,uneasy abrasion,strong user acceptance and low acquisition cost.Due to these advantages,FKP has gradually become a research hotspot in recent years.Among different FKP methods,coding-based methods are thought to be practical and important for no training,low storage and fast speed matching.The existing research on coding-based FKP recognition has the following two problems.Problem 1,most of the existing algorithms use square convolution kernel filters to extract image features,while the texture distribution of FKP has strong consistency and their line directions are relatively concentrated,which result in poor discrimination of line direction features extracted by square convolution kernel;Problem 2,in the template matching stage,almost all the existing algorithms judge the samples according to the single matching distance calculated from the extracted features between two images,while some fuzzy samples cannot be effectively distinguished by the single matching distance,resulting in high error acceptance rate and error rejection rate,which affects the overall recognition effect.Aiming at the problem that the line direction features extracted by square convolution kernel do not have good discriminative abilities in Problem 1,this paper proposes a feature extraction method based on Gabor convolution kernel transformation.In this method,rectangular convolution kernel is utilized to replace square convolution kernel to extract the features.Compared with square convolution kernel,rectangular convolution kernel can not only extract the local direction information of FKP,but also extract difference information between texture lines,so that Gabor filters with rectangular convolution kernel can extract more rich feature information;Experiments are also carried on the setting of the number of filters from effect and time to select the best number;Finally,the feature information extracted by the Gabor filter bank is encoded to obtain mean features and dominant direction features.The mean features and dominant direction features are fused to further improve the recognition effect.Aiming to release the problem that the fuzzy samples cannot be effectively judged according to the calculated single matching distance in Problem 2,a lightweight multidimensional matching distances fusion method is proposed in this paper.Based on the difference and complementarity between the matching distances of multiple codingbased algorithms,this method applies support vector machine(SVM)to classify and reprocess the constructed multi-dimensional vectors so as to effectively distinguish the fuzzy samples whose single matching distance cannot be judged accurately.Meanwhile,the proposed method is a general method and can be embedded into the existing codingbased methods.This paper verifies the effectiveness of the multi-dimensional matching distances from two to four.It also compares the performances between different dimensions and finds the best dimension,resulting in great improvement in the template matching stage.In this paper,extensive experiments based on MATLAB are carried out on the public knuckle database Poly U-FKP to prove the effectiveness of the above method.The results show that the performance of the feature extraction algorithm based on Gabor filter banks with rectangular convolution kernel is improved well,which is the best compared with existing typical coding-based algorithms;Meanwhile,compared with the single matching distance,the method based on multi-dimensional matching distances fusion has much improvement,and the effect of two-dimensional matching distances fusion is the most.It also performs best over existing typical coding-based algorithms.