CUDA accelerated iris template matching on Graphics Processing Units (GPUs)

In this paper we develop a parallelized iris template matching implementation on inexpensive Graphics Processing Units (GPUs) with Nvidia´s CUDA programming model to achieve matching rates of 44 million iris template comparisons per second without rotation invariance. With tolerance to head tilt, we achieve 4.2 million matches per second and compare our implementation to state of the art prior work performed on GPU and FPGA, emphasizing our improvements. Additionally a comparison to highly optimized CPU implementations of iris template matching is performed, showing a 14X speedup using our approach. In contrast to other published work, we develop an implementation for parallel iris template matching that incorporates iris code shifting for rotation invariance and provide timing data showing our proposed architecture is efficiently implemented, capitalizing on shared and texture memory to speedup the bit shifting process beyond current prior art.