Contactless 3D finger knuckle identification

Abstract

Biometrics is an important research area with large impact on practical applications such as immigration inspections, access control and online transactions. Compared with behavioural characteristics, physiological characteristics such as fingerprint and face are often used for reliable biometric recognition. However, every biometric identifier has its pros and cons. For example, the quality of fingerprint can be degraded by dirt and sweat, while the appearance of human face can be varied by makeup. Finger knuckle patterns on the dorsal surface of human fingers can provide discriminative information for biometric recognition. Such images can be acquired conveniently and can also be acquired simultaneously with fingerprint. On the other hand, conventional research on biometric recognition focused on the use of 2D intensity images as the source information. However, 3D images of finger knuckle contain more information which can be helpful to advance biometric recognition. Presentation attacks can also be easily detected by inspecting those 3D images. Therefore, this thesis focuses on the development of this novel biometric identifier, i.e. contactless 3D finger knuckle, which can offer an accurate, efficient, and convenient alternative for biometric recognition. The development of 3D finger knuckle recognition systems requires accurate 3D imaging for recovering discriminative knuckle surfaces. However, such 3D imaging systems are usually associated with high cost and bulk due to the nature of 3D imaging technologies. Therefore, the first stage of this thesis investigates a promising 3D reconstruction technique, i.e. photometric stereo, which requires only a single camera for recovering accurate 3D information in pixelwise-precision. This approach assumes Lambertian reflections, but most real-world objects are non-Lambertian so that the reconstruction accuracy can be degraded. In order to estimate surface normal vectors for real-world objects, we introduce two novel outlier rejection techniques which identify the reliable data for more accurate estimation of surface normal vectors. The proposed approach outperforms other state-of-the-art photometric stereo methods on a benchmark dataset of real objects. The major focus of this thesis is on developing contactless 3D finger knuckle recognition technologies. Since there are no publicly available databases of 3D finger knuckle for research and experimentation, we developed a two-session database of 3D finger knuckle images acquired from 228 subjects (with two-session from 190 different subjects). After that, we developed a contactless 3D finger knuckle recognition system with all necessary processing procedures including 3D image acquisition, pre-processing, area of interest segmentation, feature extraction and matching. We developed several new feature descriptors and matchers for accurately encoding discriminative 3D finger knuckle features and efficiently comparing pairs of feature templates. Another contribution of this thesis is on the development of deep learning based approach for this novel 3D finger knuckle recognition problem. There are several challenges in advancing such technologies, e.g. availability of very limited training data, large intra-class or train-test sample variations as observed for the real applications. We introduce a new deep neural network based approach which simultaneously encodes and incorporates deep features from multiple scales to form a more robust deep feature representation. Such collaborative feature representations are robustly matched using an efficient alignment scheme with a fully convolutional architecture to accommodate involuntary finger variations during the contactless imaging. This approach further boosts the state-of-the-art performance among other methods considered in this thesis.