Abstract

Biometric person identification refers to the recognition of a person based on the physical or behavioral traits. Palm print based biometric identification system is one of the low cost biometric systems, since the palm image can be obtained using low cost sensors, such as desktop scanners and web cameras. Because of ease of image acquisition of palm prints and identification accuracy, palm images are used in both uni- modal and multimodal biometric systems. A multi-scale and multi-directional representation is desirable to represent thick and scattered thin lines of a palm image. Multi-scale and multi-directional representation can also be used in image fusion, where two images of two different biometric traits can be fused to a single image to improve the identification accuracy. Face and palm images can be fused to keep the desired high pass information of the palm images and the low pass information of the face images. The Curvelet transform is a multi-scale and multi-directional geometric transform that provides a better representation of the objects with edges and requires a small number of curvelet coefficients to represent the curves.
In this thesis, two methods using the very desirable characteristics of the curvelet transform are proposed for both the uni-modal and bi-modal biometric systems. A palm curvelet code (PCC) for palm print based uni-modal biometric systems and a pixel-level fusion method for face and palm based bi-modal biometric systems are developed. A simple binary coding technique that represents the structural information in curvelet directional sub-bands is used to obtain the PCC. Performance of the PCC is evaluated for both identification and verification modes of a palm print based biometric system, and then, the use of PCC in hierarchical identification is investigated. In the pixel-level fusion scheme for a bi-modal system, face and palm images are fused in the curvelet transform domain using mean-mean fusion rule. Extensive experimentations are carried out on three publicly available palm databases and one face database to evaluate the performance in terms of the commonly used metrics, and it is shown that the proposed methods provide a better performance compared to other existing methods.