Fast and robust face recognition via parallelized L1 minimization

Many classic and contemporary face recognition algorithms work well on public data sets, but degrade sharply when they are used in a real recognition system. A major cause of this is the difficulty of simultaneously handling variations in illumination, image misalignment, and occlusion in the test image; while in some applications the gallery images can be well controlled, the test images are only loosely controlled. This thesis describes a conceptually simple but computationally intense face recognition system that achieves a high degree of robustness and stability to illumination variation, image misalignment, and partial occlusion, along with optimized parallel implementations. First, well registered training images taken under many illumination directions are captured using a novel projector-based acquisition system. The recognition system then uses tools from sparse representation to align a test face image to a set of frontal training images. To better handle severe occlusions, an extension to the algorithm is described that makes use of the knowledge that occluded pixels tend to be spatially correlated. Due to the use of multiple face images as features and the non-smooth nature of the optimization problems, these techniques have far greater computational requirements than techniques that extract low-dimensional features. Several custom L1 solvers are presented that achieve faster convergence on face data than general solvers. Optimized implementations for modern parallel computing architectures are investigated in order to build a system capable of performing highly accurate and robust recognition while remaining fast enough for use in access control systems. Optimized parallel implementations for contemporary CPU and GPU hardware are demonstrated to achieve near real-time face recognition for access control applications with hundreds of gallery users