Traditional methods of identity verification, such as username and password, and PIN numbers, are slowly being replaced by biometric verification methods. Hand Vein patterns are universally present for all, collected easily using a near IR camera, and is a permanent trait which will not be damaged or obscured by cuts, wounds, dirt, or diseases. It is a randotypic trait which is formed during the early phases of embryonic development and, hence, unique to everyone.[1,2] Vein structure is present beneath the skin and, hence, cannot be imitated like fingerprints. Above all, it is a contactless technology which alleviates the need for the user to touch the sensor, as is the case with fingerprint technology. This ensures hygiene, especially in hospital environments, while accessing patient records in a centralized database.

Image Acquisition:

To validate the accuracy and reliability of hand veins as a biometric, a database of seven hundred and forty images was created from seventy-four individuals belonging to both genders and various age groups. The hemoglobin in blood is sensitive to light in the wavelength range of 800 – 1100 nm and absorbs the same. Hence, the blood vessels in the superficial layer of the body, appear dark when compared to the other parts of the hand.[3] The WAT 902H is a monochrome CCD near IR camera. This camera was used in our imaging setup. To increase the sensitivity of the capturing setup, an IR filter of 880 nm wavelength is mounted in front of the camera lens. An unexposed Ektachrome color film served effectively as an IR filter. An IR source, consisting of IR LEDs, was used to illuminate the back of the hand. With our setup, we found that when the fist is clenched, the veins appear more prominent in the captured image rather than the veins in the palm. Hence, we have captured the vein pattern at the back of the hand. It was observed that age, gender, skin color, and room temperature did not play any role in the clarity of the vein image obtained.

Image processing:

The captured raw vein image is preprocessed through four steps: filtering, region of interest extraction, segmentation, and thinning.[4] The captured raw vein image has lot of unwanted details, such as hair, skin, flesh and bone structures. All these structures manifest themselves as high frequency components. Smoothing of the image was done using a nonlinear diffusion process. The region of interest was extracted using an iterative method of generating rectangles whose size varies with the particular hand's size. Segmentation was done using a histo-threshold based process. The gray level intensity values of the vein image vary at different locations of the image. Hence, different threshold values are chosen for different gray levels in the image. The gray levels available in the image are extracted from the histogram of the image. The thresholds chosen are 0, 255 and all minima extracted from the histogram. Thus, the image was adaptively segmented. The segmented vein image is skeletonized by a morphological operation called thinning. The structural element “l” is chosen from the Golay alphabet to do the “hit and miss” operation. The skeleton of the vein image, which would be 1 pixel wide, is generated. This makes pattern matching easier.

Feature Extraction:

The morphological features were obtained by extracting the bifurcation and termination points of the vein pattern. The statistical features were obtained by extracting the moments of the vein image.[5,6] The two feature sets were normalized and then fused to form a single feature vector. The acquired images of all the seventy-four subjects were processed using the above methods and the resulting feature vectors were stored in a database.

Results:

To evaluate the performance of the system, pattern matching was done on the images in the database.[7] The pattern matching was done using a backpropagation neural network. A person's image was compared with that same person's image templates and the probability of the image being rejected to be false was found. This was repeated for all the subjects in the database. The False Rejection Rate (FRR), which is the ratio of the number of instances of false rejection to the total number of instances, was calculated and found to be 0.3%. One person's image was compared with all the other peoples image templates and the probability of the image being accepted was found. This was repeated for all the subjects in the database and the number of false acceptances was found. The False Acceptance Rate (FAR) was calculated similarly to FRR. It was found to be 0.54%.

Souhel Ben Yacoub et al[8 ] have fused Face and Speech data and obtained an FAR of 1.07% and FRR of 0.25%. Miguel Ferrer et al.[9] have fused Hand Geometry and Palmprint and obtained an FAR of 1.6% and FRR of 0.44%. We have obtained comparatively good accuracy rates with a biometric system using only a single modality, namely the hand vein biometric.

The results obtained show that it is possible to obtain good performance, comparable to multimodal systems, using a multi feature based unimodal system. The modality used, the hand vein biometric by itself, is a higly unique trait that the other biometric traits used conventionally. The performance of the hand vein system is further enhanced by combining two independent feature sets.