Reconfigurable filter bank design via principal component analysis and low rank approximation

Feature extraction plays an important role in signal processing, computer vision, etc., and filter bank is used to comprehensively delineate the characteristics of input signals from different perspectives. Nonetheless, more detailed extraction of information necessitates the application of more filters with resulting increase in computational load in proportion to the number of filters. To alleviate computation load, this paper proposes a reconfigurable filter bank design that exploits symmetrical properties in coefficients of filter bank to share computations. The proposed method is based on principal component analysis that projects filter coefficients onto a more symmetrical vector space whereas low rank approximation while trading off between accuracy and computational efficiency, discards less important components to further mitigate computation load. This paper demonstrates a case study on Gabor filter bank, which is composed of 16 filter kernels. The experiments show that we reduce 68% additions and 78% multiplications in comparison with naive convolution process; and reduce 40% additions and 57% multiplications in comparison with conventional 2D-Gabor filter implementation.