Finger-vein network enhancement and segmentation

Abstract

Recently, an emerging biometric recognition based on human finger-vein patterns has received considerable attention. Due to light attenuation in imaging finger tissues, the finger-vein imagery is often seriously degraded. This makes network-based finger-vein feature representation greatly difficult in practice. In order to obtain perfect finger-vein networks, in this paper, we propose a novel scheme for venous region enhancement and finger-vein network segmentation. First, a method aimed at scattering removal, directional filtering and false vein information suppression is put forward to effectively enhance finger-vein images. Then, to achieve the high-fidelity extraction of finger-vein networks in an automated manner, a matting-based segmentation approach is presented considering the variations of veins in intensity and diameter. Extensive experiments are finally conducted to validate the proposed method.

Appendix

Let ∆ω denote the frequency bandwidth in octaves, $\triangle\varphi$ denote the half-magnitude orientation bandwidth, a _m and b _m respectively represent the short axis and the long axis of a half-magnitude profile of Gabor frequency response in mth scale, as shown in Fig. 18, the following relationships should be determined [27, 35, 36] to make half-magnitude profiles mutually tangent in the spatial frequency domain.

$$ \left\{ {\begin{array}{*{20}c} {\sigma _{m} = \sqrt {\ln 2/2} /(f_{m} \beta \pi )} \hfill \\ {f_{m} = \varsigma f_{{m - 1}} } \hfill \\ {\vartriangle \varphi \approx 2\arcsin (a_{m} /2f_{m} )} \hfill \\ \end{array} } \right., $$

(16)

where

$$ \left\{ {\begin{array}{*{20}c} {\beta = (2^{{\vartriangle \omega }} - 1)/(2^{{\vartriangle \omega }} + 1)} \hfill \\ {\varsigma = (1 + \beta )/(1 - \beta )} \hfill \\ \end{array} } \right., $$

Implementing Fourier transformation for G ^e _mk (x, y), the parameter a _m can be derived as

$$ a_m=\frac{\gamma\sqrt{2\ln2}}{\sigma_m\pi}. $$

(17)

Refer to Eq. (16), we can obtain

$$ \sigma_m f_m= \frac{1}{\beta\pi}\sqrt{\frac{\ln2}{2}}. $$

(18)

Based on Eqs. (17, 18), and Fig. 18, we can reduce

$$ \triangle\varphi \approx 2\arcsin\left(\frac{\gamma\sqrt{2\ln2}}{2\pi\sigma_mf_m}\right)=2\arcsin(\gamma\beta). $$

(19)

Let N be the number of contours with minimum redundancy in a certain scale, $\triangle\varphi=\pi/N$ is satisfying. Based on Eq. (19), the aspect ratio γ of the elliptical Gaussian envelope is approximately determined by

$$ \gamma\approx\sin\left(\frac{\pi}{2N}\right)/\beta. $$

(20)

Therefore, given four parameters ▵ω, σ ₁ (the biggest scale), M and N, a bank of even-symmetric Gabor filters with minimum redundancy can be designed accordingly.

Fig. 18

Half-magnitude profiles in the frequency domain