Image Normalization Techniques for Robust Face Recognition

Histogram equalization has been the dominant image preprocessing technique in the field of face recognition for years now. With the property of increasing the global contrast of an image while simultaneusly compensating for the illumination conditions present at the image acquisition stage, it represents a useful preprocessing step, which can ensure enhaced and more robust recognition performance. Several, more elaborate techniques, such as the multiscale retinex approach or anisotropic smoothing, have been proposed by researchers to take the place of histogram equalization, but have ulitimately been found to be more of a complement than a real substitute. However, by closer examining the characteristics of histogram equalization, one can quickly discover that it represents only a specific case of a more general concept of histogram remapping techniques (which may have similar characteristics as histogram equalization does). While histogram equalization remapps the histogram of a given facial image to a uniform distribution, the target distribution could easily be replaced by an arbitrary one. With no theoretical proof of a specific distribution being particularly suited for the task of face recognition, at least the question arises of why the uniform distribution is commonly chosen for the histogram remapping. In this paper we present an empirical assessment of the concept of histogram remapping. We report comparative results obtained on the XM2VTS and YaleB databases for four target distributions, i.e., the uniform, the normal, the lognormal and the exponential distribution, and conclude that similar or even better recognition results can be achieved when other than the uniform distribution is considered for the histogram remapping.