Reducing IO bandwidth for GPU based moment invariant classifier systems

This paper introduces an IO bandwidth reduction technique for real-time moment invariant classifier systems running on both CPUs and GPUs. This system can run in real time on commodity general purpose graphics processor unit (GPGPU) systems. The output IO is reduced by calculating the locations of objects of interest using a projection of the 2D classified outputs onto the two axes of the image. The two projections are then used to calculate the positions of a large proportion of the hits in the original image. For a system with a low number of hits there is no loss during this compression, while a system with a large number of hits only suffer losses in a small number of degenerate cases that have a low probability of occurrence in real classifier systems. Lower compression rate approaches can reduce the probability of losses at the expense of higher bandwidth and potentially lower frame rates.