Abstract
Image preprocessing techniques represent an essential part of a face recognition systems, which has a
great impact on the performance and robustness of the recognition procedure. Amongst the number of techniques
already presented in the literature, histogram equalization has emerged as the dominant preprocessing technique
and is regularly used for the task of face recognition. With the property of increasing the global contrast of the
facial image while simultaneously compensating for the illumination conditions present at the image acquisition
stage, it represents a useful preprocessing step, which can ensure enhanced and more robust recognition performance.
Even though, more elaborate normalization techniques, such as the multiscale retinex technique, isotropic
and anisotropic smoothing, have been introduced to field of face recognition, they have been found to be more of a
complement than a real substitute for histogram equalization. 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 with an arbitrary one. As there is no theoretical justification of why the uniform
distribution should be preferred to other target distributions, the question arises: how do other (non-uniform) target
distributions influence the face recognition process and are they better suited for the recognition task. To tackle this
issues, we present in this paper an empirical assessment of the concept of histogram remapping with the following
target distributions: the uniform, the normal, the lognormal and the exponential distribution. We perform comparative
experiments on the publicly available XM2VTS and YaleB databases and conclude that similar or even
better recognition results that those ensured by histogram equalization can be achieved when other (non-uniform)
target distribution are considered for the histogram remapping. This enhanced performance, however, comes at a
price, as the nonuniform distributions rely on some parameters which have to be trained or selected appropriately
to achieve the optimal performance.
Links and resources
Tags