Two directional multiple colour fusion for face recognition

Colour plays an important role in face recognition and different colour combinations can give different recognition performances. Currently, most of the existing colour component extraction approaches are based on holistic operations as well as only using three-colour components. Recently, a new approach called the Multiple Colour Fusion model (MCF) is proposed for face verification based on more than three colour components but still with holistic manipulations. In this paper, we will investigate the MCF approach for face recognition from a different perspective with local patch operations. First, with a given training-testing dataset, we can extract multiple important colours for all given local patches based on the MCF approach. In this case, the number of extracted colour components on different patches may be different. Secondly, with each prioritized colour component for each patch, we can consequently select different layers via greedy selection approach and then create a three dimensional colour map template. Finally, the resultant template is used for face recognition. With extensive experiments in AR face database, and Curtin face database, we have demonstrated that the proposed approach can improve recognition accuracy from 3.37% to 7.58% in comparison with the original MCF and other different state of the art colour extraction approaches.