Research On Robust High Precision Palmprint Recognition Technology

The edges of biometric identification like stability, uniqueness, convenience, security make us believe that it will be able to gradually displace the traditional approaches in the near future. Palmprint is one important biometric feature, and its study has attracted much attention in the past decades. Although palmprint-based recognition methods have been proposed and successfully applied for identity authentication, the image for recognition is usually natural light image captured by contact devices in the previous work. This enables the system to be attacked easily which leads to a big security hole and meanwhile it is hard for further improvement of recognition accuracy.For those issues mentioned above, this paper proposes two feasible approaches. One is that the palmprint image is captured by contactless devices, which effectively prevents the loss of palmprint template. Another is that the multispectral palmprint is used instead of natural light palmprint for recognition, which can effectively prevent the invasion of fake palmprint meanwhile further improve the recognition accuracy. However, follow-on problems are shown as follows:(1) Due to the angle and position of the capture, as well as defocus of the device when using the contactless devices, it is inevitable to have some distortions, such as translation, rotation, noise, illumination, blurriness, which would degrade the performance of recognition system. Therefore, VO image decomposition model is used to obtain stable structure features from the palmprint image and WRHOG method is designed to further extract features from the structure layer.(2) For the multispectral palmprint recognition, how to fuse palmprint features obtained from different bands effectively in order to obtain higher and more robust recognition results is what we face the great challenge, a novel hierarchical approach is proposed. First, BDOC method is designed to do a rough matching, and then a classification is done for the results from rough matching, including true match, false match, remain to be further matched. In those regions that remain to be further matched, HOG is used as the fine feature and thereby achieve a fine matching. Finally, the final matching score is obtained by integrating hierarchical matching results. Furthermore, to further improve distinguishability of features and recognition accuracy, we fuse different features from different bands in the proposed feature fusion scheme based block selection, namely choose image block with maximal response for gradient magnitude as fused image block. Extensive experiments on the PolyU palmprint databases are carried out, which obtain robust and high accurate recognition results, and validate the effectiveness and real-time of the proposed method.