Effects of measurement conditions on the projection of faces in the eigenface space: An experimental investigation

Automatic face recognition is vital for many security applications. From a scientific standpoint, it may be considered a classification process, based on the measurement of a "face", where a face is a property of a person, representable as a surface in a three-dimensional physical space. In this perspective, face measurement may be considered as derived from length measurement. The space of the measurand is theoretically an infinite dimensional one: to manage it, a two dimension projection may be considered and a suited eigenvector representation of reduced dimension may be looked for. In this context, this is called an eigenface approach. From an experimental standpoint, when operating in the field, such measurement is affected mainly by the surrounding environment and by the lack of collaboration by the investigated subject. In the current vast technical literature, the performance of different recognition procedures is investigated mainly in respect of the overall performance. By converse the pro osed approach closely follows the pattern of the metrological characterisation, viz. calibration, of a measuring instrument and consists of a detailed analysis of the effects of some critical influence quantities. We have first designed and performed an extensive experimentation aimed at identifying the most critical influence quantities, such as the dimension of the face in the image (scale), the orientation of the face, the expression and illumination conditions. Than their influence has been linearly modelled and the parameters have been evaluated form a large set of recorded images controlling the influence quantities values. Lastly, some compensation of the systematic effects has been attempted, which has proved to be useful.