A discriminative model for scale, translation and rotation invariant face recognition

There are many challenges like different illumination conditions, ageing, different poses and orientation of images, limited datasets for training, and other variational conditions associated with facial recognition and verification. SIFT is a robust and popular algorithm for facial recognition due to its invariant nature towards scale, and orientation, but it has its some limitations. This paper proposes a framework in which we modify the steps of SIFT algorithm in two ways. First, for calculating extrema using a non-maximal suppression algorithm we compare the grid in fixed patches instead of whole images, and by reducing the size of the SIFT feature descriptor. For this experiment, we are using five public databases FERET, Yale2B, M2VTS, Face 94, ORL and found an improvement in terms of accuracy with respect to the existing facial recognition system. The novelty of the proposed method is that it has less computational complexity compared to original SIFT and good accuracy compared to other state-of-the-art methods.