Skin Texture as a Source of Biometric Information

Traditional face recognition systems have achieved remarkable performances when the whole face image is available. However, recognising people from partial view of their facial image is a challenging task. Face recognition systems' performances may also be degraded due to low resolution image quality. These limitations can restrict the practicality of such systems in real-world scenarios such as surveillance, and forensic applications. Therefore, there is a need to identify people from whatever information is available and one of the possible approaches would be to use the texture information from available facial skin regions for the biometric identification of individuals.

This thesis presents the design, implementation and experimental evaluation of an automated skin-based biometric framework. The proposed system exploits the skin information from facial regions for person recognition. Such a system is applicable where only a partial view of a face is captured by imaging devices. The system automatically detects the regions of interest by using a set of facial landmarks. Four regions were investigated in this study: forehead, right cheek, left cheek, and chin. A skin purity assessment scheme determines whether the region of interest contains enough skin pixels for biometric analysis. Texture features were extracted from non-overlapping sub-regions and categorised using a number of classification schemes. To further improve the reliability of the system, the study also investigated various techniques to deal with the challenge where the face images may be acquired at different resolutions to that available at the time of enrolment or sub-regions themselves be partially occluded. The study also presented an adaptive scheme for exploiting the available information from the corrupt regions of interest.

Extensive experiments were conducted using publicly available databases to evaluate both the performance of the prototype system and the adaptive framework for different operational conditions, such as level of occlusion and mixture of different resolution skin images. Results suggest that skin information can provide useful discriminative characteristics for individual identification. The comparison analyses with state-of-the-art methods show that the proposed system achieved a promising performance.