Band Selection for Palmprint Recognition

Biometrics based on palmprint has been developing fast in the past ten years. The newly proposed hyperspectral imaging can provide high accuracy and abundant information about the palms and the tissues, veins underneath. However, due to the limitations of computation speed and storage, we have to select the most representative bands for palmprint recognition. This paper proposes a band selection scheme for hyperspectral palmprint recognition. First, the images with high image entropies and Equal Error Rate (EER) are selected. Then a clustering method is introduced to choose the most representable bands. In our experiments on the HK-PolyU Hyperspectral Palmprint Database, three bands combination can generate the best EER 0.17325%. The proposed approach can also be used for band selection of other hyperspectral systems. 


I. INTRODUCTION
Palmprint recognition has been developing rapidly in the past ten years, not only due to the development of recognition algorithms, but also the help of cutting edge capturing devices.With depth camera and multispectral/hyperspectral camera being used in palmprint recognition technology, biometrics based on palm vein [1] [2], 2D and 3D palm print [3] [4], multispectral palmprint [5] [6], and hyperspectral palmprint [7] [8] are made possible.
The human palm has abundant information both on the surface and underneath the skin.On the skin, the main lines, wrinkles and small creases are unique for each person.Underneath the skin, the pattern of the veins are special as well.
Combining different wavelength palmprint images can provide more distinctive information about the palm, and thus improve the performance of the palmprint recognition system.Ying Hao [5] used pixel-level fusion of multispectral images for the multispectral palmprint recognition and achieved best Equal Error Rate (EER) with two bands combination using Curvelet transform and shift-invariant digital wavelet transform.Dong Han [6] combined four bandwidth images, i.e.Red, Green, Blue and Infrared using wavelet transform to improve the Manuscript received December 1, 2015; revised April 11, 2016.performance.Since the hyperspectral imaging system is introduced to palmprint recognition, the amount of information can be more than enough.Zhenhua Guo [7] established a hyperspectral palmprint imaging system and used it to analyze the band selection for multispectral palmprint systems based on (2D) 2PCA method [9].His team also implemented exhaustive search for the optimal combination [8].
As we can get information more easily and more accurately, the problem of information redundancy emerges, especially in hyperspectral imaging.In hyperspectral palmprint recognition systems, it's not convenient to use exhaustive search [8] for the optimal band selection and combination.Some researchers proposed information based band selection methods [10] [11] as well.
In order to keep all the useful information about the palmprint and reduce the number of bands as many as possible, a selection of bands should be made prior to the palmprint recognition.In this paper, a band selection scheme for hyperspectral palmprint recognition is proposed.The scheme can be adjusted for other hyperspectral systems as well.First, the images captured in different wavelength illuminations are evaluated via image entropy and Equal Error Rate (EER).In the procedure, some bands containing low quality images are deleted.After that, the k-means clustering method is introduced to choose the most representable bands.In the end, a Gabor Filter based score level fusion method is proposed.

A. The Hyperspectral Palmprint Database
The hyperspectral palmprint images consist of nearly a hundred equally spaced spectral images, thus provide much more information than regular multispectral images.However, apart from information about palm prints, redundant information and even noise due to different positions of the palms are introduced to the hyperspectral palmprint images as well.
The HK-PolyU Hyperspectral Palmprint Database captured from 380 palms at 69 spectral bands with a stepwidth of 10nm over spectrum 420nm-1100nm is used for experiments [7].The Region of Interest (ROI) is cropped from the original 1004*1002 pixels CCD images to 128*128 pixels using the algorithm proposed in [12].However, due to the low transmission of Liquid Crystal Tunable Filter (LCTF) and low CCD response at long wavelength, the image quality is not consistent over the whole spectrum.

B. Spectral Band Image Entropy Method
Due to the device limitations mentioned above, the image quality varies in different spectrum.For low quality images, little information is involved and they can be discarded without affecting the recognition results.
Entropy is a measurement of the information that a grey-scale image carries by analyzing the distribution of the values of the pixels.It can also measure the distribution of the grey scales.It's defined as follows: where , is the pixel value in the 8-bit grey scale palmprint image, and m, n are the sizes of the image.
Fig. 1 shows the image entropy of each band with different wavelengths.The average entropy of the band images for 380 people are above 6 at the interval of 580nm-1030nm.While for the images captured in wavelength lower than 580nm or higher than 1030 nm, the images are unclear and only occupy a few grey scales due to the device limitations.

C. EER for Each Spectral Band
Gabor Filter is a state-of-the-art palmprint recognition method [13].A typical 2D Gabor Filter is defined as: where  = √−1, x, y is the position of the filter,  is the frequency of the sinusoidal wave,  is the orientation of the function and  is the standard deviation of the Gaussian envelope.The 2D Gabor Filter is applied to the palmprint images with six different orientations: 0,  6 Each pixel within the image will be represented with orientations, which are then coded into 3-bit.2-D Gabor Filter can extract the comparatively stable orientation information of the palm vein image.An example of the original palmprint and Gabor Filter processed palmprint image is shown in Fig. 2. The different orientations at each pixel are represented with different colors.EER is stated as a percentage and represents the point at which the false rejection rate equals to the false acceptance rate.This value indicates that the proportion of false acceptances is equal to the proportion of false rejections.The lower the equal error rate value is, the higher the accuracy of the biometric system will be.The EER computed using the Gabor Filter method can be seen in Fig. 3.The images of a single wavelength in the wavelength interval 580nm to 1080nm can contain enough information to obtain a satisfactory EER using Gabor Filter.However, for the images in the wavelength in other intervals, the EER can go up to 26% and thus introduce lots of noises or inaccuracy to the recognition system.
According to the image entropy and EER measure, it's clear that a number of bands in the middle should be selected for subsequent process.In order to select the images containing enough information and discard low quality images, the wavelength interval 580nm to 1020nm are selected for further analysis.

A. EER for Each Spectral Band
The palmprint images do not only carry information about the palmprint and palm vein, but also contain the reflection of the flash, stains on the surface and noises introduced by the capturing device.Due to different positions of the hands, the ROI images are not exactly the same, but with small rotation or offset.The Gabor filter is useful to preserve the main features of palmprint images, i.e. the main orientation in each pixel.

B. Band Selection Based on Clustering
The K-means clustering method is an unsupervised learning algorithm that can be used to reveal the relationships within the spectral bands.
The k-means method classifies a given dataset through a certain number of clusters (for example k clusters).The main idea is to define k centroids, one for each cluster.These centroids are first randomly selected from the spectral bands.The next step is to take each point belonging to a given cluster and associate it to the nearest centroid.The Hamming Distance is used instead of the commonly used Euclidean distance.Because the images are represented with orientations at each pixel, i.e.Gabor Filter values.Then the k new centroids are calculated as the center of each cluster.As a result of this procedure the k centroids change their locations step by step until the change is significantly small.Then the spectral bands are divided into k groups.

IV. EXPERIMENTS AND RESULTS
The wavelength interval 580nm to 1020nm from the Hyperspectral Palmprint Database is selected according to the image entropy and EER measure described in Section 2.
In order to evaluate the performance of the clustering method, the sum of Hamming distance between the points within the cluster and the cluster centroid is adopted.The sum is defined as S = ∑ ∑ (  () ,   )  =1  =1 (4) where (  () ,   ) is the Hamming Distance between the point   () and the cluster center   .() indicates that the th data point   () that belongs to jth cluster.The sum decreases as the number of clusters increases as shown in Fig. 4. In other words, the more samples selected, the better we approximate the images over the spectrum.
An example of band selection with cluster number k=4 is shown in Fig. 5.The wavelength interval from 580nm to 1020nm, which consists of 45 spectral bands, is divided into 4 clusters with centroids 600, 770,850 and 1000.These centroids will be selected as the representative bands for the whole spectral bands in the palmprint recognition system.In order to evaluate the performance of the selected bands, the score level fusion is used.A score level fusion based on Gabor Filter image is employed to evaluate the performance of the band selection and cluster fusion.
where   is the Hamming distance for the th selected spectral band.
The EER of the palmprint recognition system is evaluated based on the fusion score TS for different number of band selection.
The best combination according to Table I is three bands combination with EER 0.17325%.While combining more bands, not only the information adds up, but the noises will be introduced as well.The EER for 4 bands selection even drops 0.00657% compared with 3 bands selection.

V. CONCLUSION
In this paper, a band selection scheme for hyperspectral palmprint recognition is proposed.By using image entropy and EER measure, low quality bands are deleted directly.A clustering method is further used to Moreover, the biometrics system should not only have high recognition rate, but also be robust against spoof attacks.The hyperspectral images can provide abundant information to identify a real human hand, not just a printed palmprint paper or a rubber hand.In the future research, spoof attacks will be considered as well in the design of a hyperspectral palmprint recognition system.

Figure 2 .
Figure 2. Original image and the gabor filter map

Figure 5 .
Figure 5. Clustering based band selection with k=4 band combination and the clustering method is validated by the band fusion method in the end.