Computationally efficient 2-D adaptive filters based on essentially separable FIR structures

Adaptive filters are used successfully in sonar, radar, communications, seismology and biomedical engineering applications [1]. This paper presents a theory of how computational complexity of 2-D adaptive FIR filters can be reduced by using essentially separable 2-D FIR filter structures. Results illustrate that essentially separable filters are more computationally efficient than conventional 2-D adaptive filters. The trade-off is the error introduced by approximating the 2-D filter as a parallel bank of separable sub-filters. Experimental results show that the coefficients of a 2-D non-separable filter are more accurately identified as the number of separable sections increases in the essentially separable structure.