Abstract: The existing face recognition techniques cannot deal with non-uniform blurring conditions which arise from tilts and rotations in cameras. Different factors such as exposure time, stability of the platform and user experience affects the degree of blurring. In this paper a method is proposed for face recognition across motion blur, variations in illumination and pose. Images of different persons are collected and stored in the gallery. The probe set is formed using the set of above images with different blur, illumination and pose. One image among the probe set is selected and its blur is estimated using hessian matrix. A threshold value for blur is fixed. If amount of blur is greater than threshold then the estimated blur is applied to the entire gallery. The LBP (local binary pattern) features of the blurred gallery images and the given probe image is extracted and the distance between them is calculated. The gallery face which gives the minimum distance is recognised as the probe face. If the blur is less than the threshold, gaussian filter is applied to the probe image and DCT is calculated. The 3/4th of the probe image as well as all the gallery images is cropped. The HOG of LBP is found out from the cropped probe image. DCT (Discrete Cosine Transform) and HOG (Histogram of Gaussian) of LBP forms the extracted feature set. PCA (Principal Component Analysis) is used to reduce the dimension of the feature set. The feature so obtained is compared with the RDF model of all the gallery images and thus the correct user is found out.

Keywords: Blur estimation, face recognition, illumination, LBP, motion blur, pose.