Variations in illumination still pose a major constraint in face recognition systems. Though many steps have been taken in this area, it continues to be a challenging field in this domain. We propose a framework to overcome this problem by first classifying the image into dark, normal or shadowed, and then selecting an appropriate filter for the image. This step ensures that there is no loss of features in the image due to a filter that is unsuitable for the image under consideration. Also processing time is saved when normal images that do not need any filtering are skipped. The filter pre-processes the image before it can be used for any further steps such as feature extraction and matching. The illumination-classification framework is modelled on Rough Set Theory and classifies the images according to their Rough Membership Functions. The results obtained are as high as 94.28% in terms of accuracy of correct classification of images into dark, normal or shadowed. It is shown that filtering an image with an appropriate filter yields more fiducial points on a face, hence better feature extraction, and hence a stronger training system for face-matching.