In order to solve the intrapersonal variation problem in facial recognition, we propose a framework of a multi-engine system for facial recognition configurable in image types, watch sizes and engines based on performance matrices. The value of each cell in a performance matrix presents a confidence level for facial recognition; the quantified generalisation ability in a specific area of the performance matrix, corresponding to the pair of probe and training image variations, provides a reference for the selection of engines; and the cell value of a performance matrix at a specified size of watch list can be predicted through non-linear identification system approach using existing performance matrices for different sizes of watch lists. We demonstrated the improved performance of a system embedded with two engines, Eigenface and Local Binary Pattern Histograms algorithm.

Watch the webcast replay >>