Systolic architecture for hardware implementation of two-dimensional non-separable filter-bank

Abstract

In this paper, we present an efficient poly-phase decomposition scheme for implementation of 2-D non-separable filter bank. Poly-phase decomposition scheme offers multiplexing of filter bank computations or/and reduce the data clocking without affecting the overall throughput rate. Both these features can be used conveniently depending on resources availability or processor-technology. Time-multiplexing could be the choice for resource-constrained applications. Slower clocking rate could be chosen if processor-technology is the constraint. In that case, the design could be realized with cheaper and slower processor-technology. Time-multiplexed design needs proper data scheduling to perform filter bank computation interleavingly without data overlapping. Keeping this in mind, we have derived a systolic architecture for hardware realization of time-multiplexed filter bank where we have used novel data buffering scheme for the filter coefficients of the filter bank. Comparison result show that, the proposed structure involves almost J times less hardware resource than the non poly-phase filter bank structure and it provides the same throughput rate as the other, where J is the filter bank size. The hardware saving is significant for large size filter banks like Gabor. The proposed structure could be a good candidate for efficient hardware implementation of non-separable filter bank used in various image processing applications such as biometrics systems.