Research On Segmentation And Recognition Methods In Vascular Images

Compared with traditional authentication verification technologies, biometrics ensures security and convenience, and has aroused much attention from academic and business circles. Patterns of vasculars hide beneath the skins, are characterized by their live-identification, security and anti-counterfeitness. The vascular patterns mainly contain vein patterns from hand and retinal vasculatures.In this thesis, we mainly focus on finger vein recognition and segmentaion and rencogiton in retinal images.Among the existing biometrics, finger vein is generally considered as one of the most promising patterns.Comparing to other biometrics, finger vein possesses the merits of high security and offer better user experiences. Nowadays, personal verification based on biometrics is widely used in door access control, security systems, forensics and so forth.Though much progress has been made in the research of finger vein recgoniton, problems still exists. Identification based on finger vein remains a challenging task due to image quality issues and deformations.For example, the image quality and deformation problem can bring bad influence to finger vein recognition. Improving the existing feature extraction method or designing robust features is one solution to these challenges.The main structures of fundus image are blood vasculatures and retinal pigment epithelium. Retinal vasculature is the only non-traumatically ovserved part of the human circulation system. Retinal vasculature is distinctive enough to be a new kind of boimetric patterns, the measurement of various vasculature features can also throw light on early prevention of system disease and retinopathy. One of the optimal tasks is vasculature segmentation, which is a prerequisite step of both retinal recognition and pathology analysis. Various segmentaion works exits, but can’t balance the efficiency and accuracy simultaneously. How to segment the retinal vasculature fast and accurately is still a challenge. On the other hand, improperly segmented vasculatures can also bring uncertainties to retinal recognition, how to implement recognition with vascular segmentaion is worthy of exploring.Based on in-depth analysis of the problem exists in finger vein recognition and segmention and recognition of retinal images, this thesis proposes several techniques to solve the above-mentioned problems to promote recognition accuracy of finger veion and retinal recognition systems.In addition, boost the efficency and accuracy of retinal vasculature segmention tasks. The main works and contributions are as follows:1. Finger vein recognition methods based on local patterns have always been important. This kind of method mainly includes local binary pattern and its variants. It can be figured out that the existing local pattern based method only considered the gradient orientation of pixels, while the gradeint magnitude and relations of neighbor gradient are ignored. How to utilize the gradient information is a key problem to improve the accuracy of local pattern based finger vein recogntion. This thesis proposes a new local pattern called Local Direction Code (LDC), it encodes both the orientation magnitude and relations of crossover directions. Evaluated on finger vein recogniton, experiments on dataset of 4,080 images from 136 fingers demonstrate the discirminability of the prposed feature, the Equal Error Rate (EER) is 50% lower than the Local Line Binary Pattern (LLBP) based recogntion.2. When dealing with deformation problem, most existing methods regard it as a kind of noise and trying to reduce the bad influence of deformation or rectify it, while neglected the discriminant information inherited in deforamtions. We noticed analytically that through genuine matching, though deformations exist, the displacment directions and distances of pixels in near locality are similar due to constraints of positional relationship. On the contrary, it is not the case in imposter matching. Such information is viable for genuine and imposter matching, how to utilize it to overcome deformation problem and improve the recognition accuracy is worth exploring. Thus, a finger vein recognition method based on deformation information is proposed, and the uniformity of displacement matrix is used as similarity score to distinguish imposter and genuine matching. The displacement matrice are generated by optimized matching on pixel based features. Tested on two publicly available databases PolyU and SDU-MLA, the average EERs of six-fold cross-validation are 0.0010 and 0.0049, respectively, which demonstrate the discriminability of deformations and effiency of the corresponding recogniton method.3. The existing techniques for vasculature segmentation in 2D retinal images can be roughly divided into supervised and unsupervised methods.The supervised methods usually have superior performance than unsupervised methods, but the redundancy of pixel based features always aggravates time and memory consumption. However, the unsupervised methods encode the human knowledge to identify the vasculature, the results are often unsatisfactory. In addition, the two kinds of methods can’t keep the true edge of vasculature elegantly. How to segment the vasculatures effiently and improve the accuracy is a key problem. In this thesis, we propose an unsupervised retinal vasculature segmentation method based on matched filters. First, two matched filters are designed for image detail enhancement and intensity normalization, respectively. Then, the corresponding results were combined by different weights, the vasculatures can be easiy seperated by a simple threshold. After simple finetune and postprocessing, results on publicly availiable DRIVE and STARE database, the accuracies are superior to all of the unsupervised methods and most of the supervised methods, with a higher efficiency.4. The Scale Invariant Feature Transform (SIFT) descriptor is famous for its powerful discriminability, and has been widely applied in object detection and classification. However, when applied to images contain vasculature strucutures, the recognition rate is unexpectedly unsatisfactory with an EER of 0.0436 on the VARIA database. We think finding reason of why the result is undesiralbe is fundermatal for improving the efficiency of SIFT based retinal recogntion. In-depth analysis shows that it relates to the blurred and low contrast image quality problems, which has unperceptable image details and reslut in unstable keypoints. Based on the above analysis, we propose an image enhancment method based on bias removal. Moreover, we further smooth the result image by Iterated Spatial Anisotropic Smooth method. The EERs are 0 and 0.0065, respectively, on the VAIRA and secondary database, which demonstrate the robousty and effectiveness of the proposed enhancement technique has improved the accuracy of the finger vein recgontion system.