Research On Preprocessing And Matching Technologies Of Finger-Vein Image

Human finger veins are blood vessels inside human fingers and exhibit excellent advantages,such as uniqueness among individuals,permanent and clear patterns,invisibility to human eyes,dependence on living body,high-security level,and so on.Owing to these features,the finger vein recognition system has been gradually applied to various fields,including financial,automobile,and building,ever since its first application for personal identification in 2004.Particularly,in such a recognition system,finger vein images are first captured using the property of hemoglobin in blood vessels that it absorbs near-infrared light-emitting diode(LED)light;then,finger vein patterns are obtained via analyzing and processing these images;finally,authentication process is conducted based on obtained vein patterns.A typical finger vein recognition system includes four steps,including image capturing,image preprocessing,feature extraction,and image matching.Among them,preprocessing and matching are critical to achieving a high-efficiency recognition system.Preprocessing aims to improve the quality of captured images,including denoising images without losing finger vein information,enhancing vein texture,segmenting object from background,and so on.On the other hand,matching techniques seek the most similar target image through analyzing image features.In recent years,various efforts have been directed on these two fields,bringing in some progresses.However,in practical applications,captured images are usually seriously degraded by noises,like blurred vein texture,low contrast,and less feature points.This subsequently brings informidable challenges for both image preprocessing and matching.Therefore,this work focuses on investigating preprocessing(including denoise,enhancement,and segmentation)and matching techniques for finger vein images.The main contributions are summarized as follows:1.Finger vein image denoising based on sparse representation(SR)and dictionary learning(DL)The basic idea behind sparse representation is to represent a test image by a sparse linear combination of training samples(atoms)from a dictionary.This process realizes the separation between useful information and noise components,hence gets the denoised image.In previous studies,the original training images are set as dictionary atoms directly.This treatment ignores the fact that training images may vary in their effectiveness as feature descriptors,thus jeopardizing the denoising and recognition performances.To overcome this limitation,the present study proposed a new SR coupled DL for finger vein image denoising.In the proposed algorithm,based on contributions to the sparse representation result,all the atoms are classified into a more discriminative part and a less discriminative part.Then,through iterative learning and updating the dictionary,the weight for the more discriminative part is enhanced while that for the less discriminative part is suppressed.This subsequently leads to denoised images with improved signal to noise ratio.Experimental results indicate the developed denoising method can efficiently suppress image noises without losing useful finger vein information.2.Finger vein image enhancement based on Pulse Coupled Neural Network(PCNN)Due to the characteristics of nonlinear modulation and automatic threshold,PCNN has become an attractive method for finger vein image enhancement.However,it is hard for PCNN to the real-time and intelligent requirements of the finger vein recognition system,owing to the complex structure and the complicated parameter setting.Thus,a new model based on PCNN is proposed to enhance finger vein image quality.Specifically,a simplified PCNN structure is introduced to reduce computational complexity at first.Then,according to the vein direction,gray intensity,and firing conditions,parameters of PCNN are automatically determined without involving any empirical correlations or trainings,to improve the adaptability of finger vein recognition system.Compared with the existing methods,the proposed model significantly improves the quality of finger vein images as well as the recognition accuracy.3.Finger vein image segmentation based on Graph Cut(GC)methodThe Graph Cut model maps an image to a weighted graph,and designs the energy function based on the similarity among image pixels.By minimizing the predefined energy function,GC can separate the object from the background.This segmentation method however demands users’ interactions,which is laborious and time-consuming and thus cannot meet the intelligent requirement of finger vein authentication system.To overcome this disadvantage,this study proposes an iterated Graph Cut(IGC)method for automatic and accurate segmentation of finger vein images.First,geometric structures of both the image-acquisition system and fingers are applied to automatically set the hard and shape constraints of IGC.Second,according to the gray similarity among image pixels,a node-merging scheme is applied to reduce the computational burden.Then,the initial labels are automatically determined by using the Gaussian probability model.Finally,the maximum a posteriori Markov random field framework is applied which iteratively updates the object and the background labels.Experimental results indicate that the proposed IGC method can provide automatic,fast,and accurate segmentations of finger vein images.4.Finger vein image matching based on phase-only correlation(POC)The POC method calculates the matching degree between two images based on the frequency domain correlation of the Fourier power spectrum.Sharper and larger correlation values represent higher matching accuracy.But traditional POC considers spectrum and scale windows with large ranges,which causes significant computing redundancy.A modified band-limited POC(MBPOC)is proposed to reduce the cost of computation.By considering useful finger vein information locates in the low frequency domain,we define the low frequency part of the finger vein image spectrum distribution as the band-limited range.Then based on the relationship between correlation coefficient and finger vein image displacement,the displacement range between two images is chosen as the scanning range.The results show the proposed MBPOC algorithm improves the recognition speed as high as 59.6%.5.Finger vein image matching based on Convolutional Neural Network(CNN)CNN has seen a recent surge of interestand exhibited remarkable advantages in the field of image matching.However,previously CNNs have not considered the characteristics of finger vein database,thereby leaving space for model improving.In this work,experiments are carried out to analyse effects of network structures and parameters on the matching accuracy.With reference to these results,a modified CNN is thus established to realize finger vein image matching.In the meantime,due to the limitation that finger vein images usually contain less feature points,the multi-directional Gabor filter image of the finger vein image is selected as the input of the modified CNN.This further enriches the feature information and boosts the matching accuracy of the novel CNN.So,the Gabor wavelet features of finger vein image in multi-directions are calculated and introduced into the CNN as inputs,which rich the feature information and improve the matching accuracy.The results show that,the proposed improved CNN structure can achieve matching accuracy as high as 99.06%.In addition,to verify the integration feasibility of the above work in practical applications,we build two smart lock samples for finger vein recognition experiments.Compared with related finger vein recognition products,the two systems in this study can produce higher recognition performance.