Finger Vein Recognition by Combining Anisotropic Diffusion and a New Feature Extraction Method

Biometric recognition systems are becoming more important in the world and are being implemented in companies where security is at the forefront. Despite the fact that fingerprint, iris and face recognition are used for security, monitoring and access control, finger vein recognition systems have distinct advantages over other biometric recognition systems; User-friendly, vivid tissue detection, high reliability and little equipment space requirement [1]. For example, although the fingerprint is the most widely used biometric identification method, its presence on the outer surface of the skin creates safety gaps and can be easily copied. Iris and retina recognition systems come to the fore in terms of reliability. In practice, the difficulties that people have shown in bringing their eyes closer to a light source have been the biggest obstacle in spreading this recognition system. Facial recognition systems have difficulties in terms of sustainability, noise and reliability. As the vein recognition system eliminates all these difficulties, it has been used as an additional security measure in order to increase the reliability of the security methods used by the major financial sectors and banks in various countries of the world [2-4]. For example, in many countries, vein recognition systems have become more popular than fingerprint recognition systems, and vein recognition has become a preferred security method [5].

FV images are obtained using near infrared (NIR) wavelength rays. According to researches, it is determined that NIR rays can pass through human skin and absorbed by hemoglobin in the vessel at wavelengths between 740-960 nm [6]. Therefore, the reflection of the infrared ray is less in the tissue where the veins are dense and therefore the regions where the veins are dense appear dark. This can be visualized using near infrared sensitive Charge Couple Device (CCD) cameras [6]. In a study by Kumar and Zhou, it was shown that the FV images are different even in identical twins and the vascular network of each finger is different from each other [7].

However, despite all these distinctive features of the FV, there are some difficulties in the use of FV images. Light scattering in the tissue and thus image defect is one of the main problems that occur when FV image is taken by CCD cameras [8]. This light scattering problem significantly affects the visibility of the vessels as poorly and reduces the success of classification [9]. For these reasons, in this study, Homomorphic Filter and Perona-Malik Diffusion, have been applied in order to get rid of the light scattering in our captured image, to increase the image quality and the visibility of the vein regions and to prepare our image for segmentation step. In addition, two new features have been proposed based on the ratio of the distance between the veins intersecting the horizontal and vertical lines passing through the center of gravity of the FV images in order to improve the classification success in FV recognition.

The rest of this paper is organized as follows. Section 2 provides information about some of the studies in the literature for FV recognition. Information about the main methods used in this study can be found in Section 3. In Section 4, information can be found about the flowchart of our algorithm and the definitions of the new attributes we propose. The results of our study and comparisons of similar studies in the literature are explained in Section 5. Finally, the conclusion of this study is shown in Section 6.

3.1 Finger vein image database

Yang and Zhang [30] was asked to submit a database of FV images. The finger vein database consists of 15 index finger images of 64 persons. The dimensions of the samples in our database are 170x80 pixels on average and *.jpg format. Our database consists of 960 images in total. FV images were obtained via NIR sensitive CCD camera using NIR light source at 760 nm wavelength [28]. Figure 1 shows a sample FV image from the database.

1.png

Figure 1. Original FV image

3.2 Homomorphic filter

Homomorphic Filters are widely used in digital image processing to enhancement and restoration to get rid of local imbalances in exposure. This exposure imbalances usually occur when a part of an object is strongly illuminated while the other part of the object is slightly illuminated. The difference between the illumination can depends on the property of the object under the light source. Since biological tissues have a high degree of inhomogeneous optical environment, there may be lots of light scattering in the images obtained with these entering rays [14].

Homomorphic Filter is based on the theory that the light distribution in a digital image is defined by a multiplication of the reflectance of the objects and the scene illumination as described by:

$Image(x, y)=\operatorname{Light}(x, y) *RefObj (x, y)$     (1)

where, Light(x,y) is the intensity of the illuminating Light source and RefObj(x,y) is the reflectance of the object between 0 and 1. Transforming the image into the log domain, this can be expressed in the density domain as

$\ln (I)=\ln (L)+\ln (\rho)$     (2)

thereby creating a new image description, $\ln (L)$ is low pass and $\ln (\rho)$ is high pass and we have an $\operatorname{Image} I=\ln (I)$ which has additive noise can be selectively filtered by a linear filter. Since the illumination is assumed to have low-pass characteristics, the filter needs to be high pass with sufficiently large spatial support to cover the variations of the illuminating lightsource [31]. This leads to the common implementation of the above equation taking a Fourier Transform, applying High Pass Filter and taking the inverse Fourier Transform It has effectively divided the lightsource contributions out of the image. The certain contributions of the lightsource are still in the image, whereas certain contributions of the object are no longer present, based upon the quality of the filter choice. Figure 2 shows the operation steps, input and output pictures of the Homomorphic Filter.

2_2.png

Figure 2. Flowchart of homomorphic filter

3.3 Perona-Malik diffusion

Smoothing is the result of a diffusion process. Diffusion processes derive from Fick’s law [32] is a physical process which balances the concentration difference without destroying the data. So that smoothing process can be considered as a diffusion process. Perona-Malik Diffusion, which is an Anisotropic Diffusion, is the space variant smoothing filter depending on the data content [33, 34]. This technique reduces the high frequency components, without removing significant parts of the data content.

Diffusion equation can be considered as [34]:

$\frac{\partial u}{\partial t}=\frac{\partial}{\partial x}\left(C \cdot \frac{\partial}{\partial x} u\right)$     (3)

where, C is the diffusion coefficient, u is the input data. If the value of C=1 then the result will be the same as the Linear Diffusion filter, and Diffusion Equation will be:

$\frac{\partial u}{\partial t}=\frac{\partial^{2}}{\partial x^{2}} u$     (4)

The idea of Perona-Malik is making the diffusivity signal dependent. The value of C must be between 0 and 1. When the process is near significant parts of the data, the value of C converges to 0 in the other parts of the data it converges to 1. The value of C is determined in Eq. (5) by using the magnitude of the first derivative

$C=g\left(\frac{\partial}{\partial x} u\right)^{2}$     (5)

by using Eq. (5) in Eq. (3), anisotropic Perona-Malik Diffusion Equation [35] is obtained in Eq. (6):

$\frac{\partial u}{\partial t}=\frac{\partial}{\partial x}\left(g\left(\frac{\partial}{\partial x} u\right)^{2}\right)$      (6)

The result is obtained in Eq. (7):

$g\left(\frac{\partial}{\partial x} u\right)=\frac{1}{\left(1+\left(\frac{\left(\frac{\partial}{\partial x} u\right)^{2}}{\lambda^{2}}\right)\right)}$     (7)

Lambda (λ) term in Perona-Malik equation can be thought as a contrast parameter separating regions of forward diffusion from regions of backward diffusion.

3.4 Niblack segmentation

The FV pictures have some contrast and light scattering problem. That’s why multi-thresholding methods is needed. Niblack’s Segmentation Algorithm [36] is one of the locally adaptive thresholding methods. In this algorithm, a kind of multi-thresholding segmentation method is used and threshold calculates at each pixel and this depends on the local variance. This method calculates the local average and local standard deviation by using Eq. (8).

$\mathrm{T}(\mathrm{i}, \mathrm{j})=\mathrm{m}(\mathrm{i}, \mathrm{j})+k \cdot \sigma(\mathrm{i}, \mathrm{j})$     (8)

In the Eq. (8), m(i,j) is the average value of the number of pixels, σ(i,j) is the standard deviation in a fixed size (i,j) window and k is a constant value which will be chosen by the user. The size of the window should be chosen as small as possible to preserve the details in the picture. At the same time it has to be chosen big enough to suppress the noise.

The finger vein database consists of 15 index finger images of 64 persons. This database consists of 960 images in total. FV images were obtained via NIR sensitive CCD camera using NIR light source at 760 nm wavelength [28]. In this section, two different experimental studies have been carried out using spatial domain and frequency domain features. The k-fold cross validation technique was used for training and testing of the database. The considered k values for the cross validation are k=3 and 5.

5.1 Experiment 1

Using the Flowchart shown in Figure 3, feature vector was created by obtaining the center of gravity (x, y coordinates), Entropy, Standard Deviation, HTP and VTP features by 1x6 dimension for each FV image in the database. Depending on the quality and quantity of some metrics, the performance of different classifiers will change [37]. Complex Tree, Ensemble, SVM and KNN were used as classifiers in this study. For Complex Tree, 100 is defined as Maximum number of splits and Gini’s Diversity Index is used as Split criterion. In the Ensemble classifier, AdaBoost, Subspace, RUSBoost and Bag were used as methods, and the highest success was achieved with Bagged Trees. 1-to-1 multiclass classifier method was preferred and Linear Kernel Function was used for SVM. Different k values were tried for KNN and the best result was obtained for k=1. The performance of different distance metrics (Euclidean Distance, Mahalanobis Distance and City Block Distance) for KNN were examined and the highest classification accuracy was obtained by using City Block Distance. The k-fold cross validation technique was used for training and testing (k=3 and k=5). The results obtained are shown in Table 2.

In this study, in order to see the effects of the proposed HTP and VTP features on the classification performance, the HTP and VTP features were subtracted from the feature vector and the classification process was performed again. Classification accuracy for 5-fold cross validation is superior to 3-fold cross validation.

As shown in Table 2, in this experimental study, the classification performance of all classifiers increased when HTP and VTP were used. Thus, it can be said that the proposed HVTP features cause an increase in performance in all classifiers we use in the study.

5.2 Experiment 2

After the preprocessing step in the flowchart shown in Figure 3, 1x52 dimension feature vector was extracted for each image using Generic Fourier Descriptor [38] and Classification of FV images was performed using these features. The same classifiers and k-fold cross validation technique (k = 3 and k=5) used in Experimental 1 were used for classification. The best result for the SVM classifier was obtained using the Linear Kernel Function. The results obtained are presented in Table 2.

Table 2. Comparison of classification performance for different features and classifiers

Table 3. Comparison of training time and prediction speed for classifiers

Table 4. Results of the studies in the literature and the result of the proposed method

In this study, in order to see the effects of the proposed HTP and VTP features on the classification performance, the classification process was re-performed by using HTP and VTP features together with the Generic Fourier Descriptors. As shown in Table 2, in this experimental study, the classification performance of all classifiers increased when HTP and VTP were used. For SVM and KNN classifiers, the success achieved by 5-fold cross validation is superior to 3-fold cross validation, but not by Complex Tree and Ensemble classifiers. The highest classification performance was obtained by using the KNN classifier. According to the results in Table 2, for both k values in k-fold cross validation technique, it was observed that HVTP features significantly increased the classification performance of Complex Tree and Ensemble classifiers.

Considering all experimental studies, the highest classification performance was obtained by the KNN classifier when Generic Fourier Descriptor and HVTP features were used together.

Comparison of training time and prediction speed for classifiers is shown in Table 3. Since the number of features has increased when HVTP attributes are added, the increase in training time and the decrease in prediction speed is quite normal. The decrease in prediction speed is due to the increase in test time, which increases in parallel with the increase in training time. Considering the increase in the success rate, it is seen that the increase in time is acceptable level as shown in Table 3. Among the classifiers used in this study, the KNN classifier gave the best results in terms of both time advantage and success rate.

Different datasets generally have been used in the studies in the literature. In these studies, features, classifiers, the number of samples and the number of classes are vary. Table 4 gives information about the studies performed in the literature and shows the results obtained [26-28, 39-41]. At the same time, the results obtained in our study were compared with the studies in the literature. The proposed method was found to produce acceptable results.