The document proposes a method for face recognition in the presence of non-uniform motion blur from hand-held cameras. It models a blurred face as a convex combination of geometrically transformed gallery images. It develops an algorithm using the assumption of sparse camera motion and an l1-norm constraint. The framework is extended to handle illumination variations by exploiting the bi-convex set of images from blurring and illumination changes. The method is also extended to account for pose variations and uses a multi-scale implementation for efficient computation and memory usage.