Study Of Finger Vein Recognition Methods

Finger vein recognition is a new physiological characteristic based biometric technique, and it uses the vein pattern in finger palmer for human identification. Same as other biometric traits, finger vein recognition has four processing steps, i.e., image capturing, image preprocessing, feature extraction and matching. In each step, there are some urgent problems. For example, as the random finger placement in image capturing, some rotated images appear, but this problem does not receive proper attention. And the existing region of interest extraction methods are designed for images captured by one specifical device, which causes the poor performace of these methods in processing images from other devices. The second case is that, the disconnected and noisy vein pattern and ineffective matching open a room for the vein pattern based methods. Assisting primary trait by the soft trait is an effective way for enhancing the recognition performance in biometrics, but there is no related work in finger vein recognition, which is the third case. Lastly, affected by personal difference and unsatisfactory performance of acquisition device, there are some low quality finger vein images, so how to evaluate the image quality is also very important problem in finger vein recognition.This thesis focuses on four mentioned problems, and main works are showen in the following:(1) A sliding window-based region of interest extraction method is proposed. Firstly, the center line of finger boundaries is calculated, and employed to detect and correct the rotated image. Secondly, for the corrected image, the interohalangeal joint in finger is located by a sliding window, and the position of the joint is further used to ascertain the height of region of interest (ROI). Lastly, the ROI image is obtained by using the internal tangents of finger boundaries to ascertain its width.(2) A superpixel based corss-sensor ROI extraction method is proposed. The images from multiple sensors are varied in image size, gray level of image backgrounding and noises, which makes the performace of the exsiting single-sensor method non-ideal. But there are still two common points in these images:A) Noises are always located outside the finger area; B) There is big gray difference on the finger boundary. So, the superixel over-segmentation is used, in which the finger area, image background and noises can be grouped into different image patches. With this in mind, a superpixel based ROI extraction method is proposed. In detail, the finger boundaries are tracked from the boundaries of superpixels, and the finger boundaries are further used in ROI location.(3) A novel finger vein recognition framework is proposed, including an anatomy structure analysis based vein extraction algorithm and an integration matching strategy. One limitation of the existing vein pattern based methods is lacking deep analysis of vein anatomy characteristic and imaging characteristic. In the proposed framework, based on the valley or half valley shaped cross-sectional profile of vein pattern and the characteristics of vein anatomy structure, anatomy structure analysis based algorithm is employed to extract vein network and vein backbone. In matching, vein backbone based image calibration is used to overcome the large scale finger displacement problem, and the elastic matching is used to alleviate the small scale vein deformation problem. And the vein backbone and vein network are integrated in matching to representated finger vein image from both the macro view and micro view.(4) A new soft biometric trait is explored in finger vein recognition. In this work, the width of the distal interohalangeal joint is measured and used as a new soft trait to assist the primary trait in finger vein recognition. And in order to effectively fuse the soft trait and primary trait, three framwroks are proposed, i.e., fusion framwork, filter framework and hybrid framework.(5) A support vector machine based finger vein image quality evaluation method is developped. To comprehensively represent image quality, three quality features are used, i.e., gradient, image contrast and information capacity. Considering the quality classification is a non-linear and small sample problem, support vector machine is used as classifier. And owing to the number of high quality images is larger than it of low quality ones, R-SMOTE algorithm is employed to overcome the class imbalance problem.