Multimodal biometrics fusion
Biometric is one of the most promising means to prevail conventional security methods. The main reason being that biometric gives authority directly to the user and not to indirect means such as a token or a password. However, there remain many problems to be resolved before biometric can really gain wide applications. For example, due to inherent limitations as well as external sensing factors, no single biometric method can warrant a hundred percent authentication accuracy as well as universality of usage by itself without imposing strict conditions during usage. Noting that these problems may be alleviated by combining multiple biometric methods, we are thus motivated to seek effective means to combine several biometrics to improve the situation.
Face and hand biometrics
The face and the hand are among the most conveniently accessible human body parts. This possibly explains the pervasive usage of face and fingerprint in authentication applications, particularly those related to personal identity and travel documents. Due to the large intra-class variations which overlap much into the inter-class variations, as well as the inherently large image size which leads to high dimensionality, face recognition based on 2D images is considered to be among the most challenging problems. Likewise, identity verification based on 2D fingerprint images is also a challenging task even though the images could be acquired under a much controlled environment. Our main focus in these themes will be on deriving effective means for feature extraction.
Apart from using a biometric for authentication, the security of the biometric itself remains an issue. Under this research theme, we will focus on revocable biometrics where a feature template can be easily replaced when it is compromised.
Pattern Classification and Machine Learning
Pattern classification is a key component in many decision processes spanning various fields of research in science and engineering. Examples include biometric authentication, medical diagnosis, information fusion and data mining. Apart from statistical means, learning from examples constitutes a major paradigm in pattern classification. Under this paradigm, a classifier minimizes a cost function and forms a map between the input feature space and the output hypothesis space of possible functional attributes such as pattern classes. In view of the nonlinearity incurred from the error counting process thereby requiring an almost inevitable iterative search for an optimized solution, we seek a deterministic approach in this work.
Recently, we have found that a zero-one counting step in the minimum classification error (MCE) problem can be effectively approximated by a quadratic function, and this approximation together with a linear parametric learning model give rise to a closed-form solution minimizing the MCE. The approximation has also been extended to maximize the area under ROC for biometrics fusion where, again, a closed-form solution can be obtained.
Network learning and optimization
As an universal approximator, the artificial neural network becomes a natural choice in many black box learning applications. For a feedforward network with random hidden neurons, the above mentioned quadratic approximation holds well for the zero-one loss objective used in training and can be well adapted with good classification property. Moreover, through a dual formulation of the solution, the learning can always be either determined or over-determined.