Palm Vein Recognition Based on Multi-algorithm and Score-Level Fusion

Abstract

In order to improve the recognition rate of palm vein recognition algorithm, a recognition algorithm based on SIFT and ORB features extraction and score-level fusion is presented in this paper. Score-level fusion is an information fusion technique, which has six common combination rules. Two dynamic weight combination rules are proposed as a supplement here. The main steps of the proposed algorithm are: First, extract Region of interest (ROI) from the registered palm vein image and the to-be-matched palm vein image and process them with sharpen enhancement, and then extract SIFT features and ORB features and obtain matching scores respectively, finally utilize score-level fusion to compute the final score for decision. The experiments on the CASIA Palm Vein Image Database show that the algorithm attains the best recognition rate by utilizing the min-rule, and the equal error rate (EER) is 0.36%.

A Novel Image Enhancement Method for Palm Vein Images

“…Our experiments with separate left/right hands matching indicated that reducing the total number of inter class matches causes a significant drop of EER values. Our experiments with bidirectional matching methods on SIFT [9] and RootSIFT [8] resulted in higher EER values than the original research. Especially the EER value difference is significantly high against [9].…”

“…To assess the proposed image enhancement method, several palm-vein image recognition systems which use SIFT and RootSIFT features with the CASIA dataset have been implemented. With the SIFT features we applied Euclidian Distance (ED) [6] [8] [9], ED then RANSAC filtering [7].…”

“…In [6], ECS-LBP is applied prior to extracting and matching SIFT features using ED and matching between left and right hands was separated. In [9] a Prewitt high pass filter was used to enhance the images before extracting SIFT and ORB features. Features are then matched using ED bidirectionally between the template and probe images.…”

“…This method is referred to as Multiple Overlapping Tiles (MOT). We further test the performance of the proposed enhancement method using existing SIFT and RootSIFT based palm vein recognition systems [6]- [9]. In section III Fig.…”

“…These LBP based image enhancement methods clip out and discard most of the subspace information which makes them unsuitable to use with invariant features. Yan et al [9] discussed an image sharpening technique which they first multiply the image with an enhancement factor of 10, and then filtering the resulting image with the Prewitt high pass operator. This is then subtracted from the original image.…”