Face recognition using skin texture

In today's society where information technology is depended upon throughout homes, educational establishments and workplaces the challenge of identity management is ever growing. Advancements in image processing and biometric feature based identification have provided a means for computer software to accurately identify individuals from increasingly vast databases of users. In the quest to improve the performance of such systems in varying environmental conditions skin texture is here proposed as a biometric feature.

This thesis presents and discusses a hypothesis for the use of facial skin texture regions taken from 2-dimensional photographs to accurately identify individuals using three classifiers (neural network, support vector machine and linear discriminant). Gabor wavelet filters are primarily used for feature extraction and arc supported in later chapters by the grey-level cooccurrence probability matrix (GLCP) to strengthen the system by providing supplementary high-frequency features. Various fusion techniques for combining these features are presented and their perfonnance is compared including both score and feature fusion and various permutations of each.

Based on preliminary results from the BioSecure Multimodal Database (BMDB) , the work presented indicates that isolated texture regions of the human face taken from under the eye may provide sufficient information to discriminately identify an individual with an equal error rate (EER) of under 1% when operating in greyscale.

An analysis of the performance of the algorithm against image resolution investigates the systems performance when faced with lower resolution training images and discusses optimal resolutions for classifier training. The system also shows a good degree of robustness when the probe image resolution is reduced indicating that the algorithm provides some level of scale invariance. Scope for future work is laid out and a review of the evaluation is also presented.