SVD-Based Minimax Design of 2-D Digital Filters
Tian–Bo Deng
Abstract
Tian–Bo Deng
Abstract
Two-dimensional (2-D) digital filters have found useful in image processing and other 2-D signal processing fields. As compared with one-dimensional (1-D) digital filter design, 2-D filter is more difficult to design because it requires more complicated mathematical formulations. Also, as the order of 2-D filter increases, the 2-D filter complexity increases drastically. One of the most efficient ways to design a 2-D filter is to decompose the difficult 2-D filter design problem into a set of 1-D filter design problems, which are relatively easy to solve. Through solving the 1-D filter design problems, a 2-D filter can be indirectly obtained. Moreover, exploiting symmetries can further reduce the complexity of the resulting 2-D filter [1]. This paper presents a special structure for designing low-complexity two-dimensional (2-D) half-band digital filter with very sharp transition band. Performing the singular-value-decomposition (SVD) of zero-phase half-band frequency response generates a set of one-dimensional (1-D) frequency responses which can be approximated using the cascade of a half-advance digital filter and a set of 1-D filters. Each such a 1-D filter only requires a single multiplication to get an output sample, so the resulting whole 2-D half-band filter has considerably reduced computational complexity as compared with the existing direct design approach. To minimize the peak errors of the magnitude response of the resulting 2-D half-band filter, we parallelize one extra section to the structure and then minimize the peak errors using a nonlinear optimization method, which reduces the design problem to a minimax design. Since the whole structure only consists of a set of simple 1-D filters, the complexity is extremely low.
A significance statement is not available in the OpenAlex record.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Two-dimensional (2-D) digital filters have found useful in image processing and other 2-D signal processing fields. As compared with one-dimensional (1-D) digital filter design, 2-D filter is more difficult to design because it requires more complicated mathematical formulations. Also, as the order of 2-D filter increases, the 2-D filter complexity increases drastically. One of the most efficient ways to design a 2-D filter is to decompose the difficult 2-D filter design problem into a set of 1-D filter design problems, which are relatively easy to solve. Through solving the 1-D filter design problems, a 2-D filter can be indirectly obtained. Moreover, exploiting symmetries can further reduce the complexity of the resulting 2-D filter [1]. This paper presents a special structure for designing low-complexity two-dimensional (2-D) half-band digital filter with very sharp transition band. Performing the singular-value-decomposition (SVD) of zero-phase half-band frequency response generates a set of one-dimensional (1-D) frequency responses which can be approximated using the cascade of a half-advance digital filter and a set of 1-D filters. Each such a 1-D filter only requires a single multiplication to get an output sample, so the resulting whole 2-D half-band filter has considerably reduced computational complexity as compared with the existing direct design approach. To minimize the peak errors of the magnitude response of the resulting 2-D half-band filter, we parallelize one extra section to the structure and then minimize the peak errors using a nonlinear optimization method, which reduces the design problem to a minimax design. Since the whole structure only consists of a set of simple 1-D filters, the complexity is extremely low.
Key concepts: Filter design, Half-band filter, m-derived filter, Filter (signal processing), Root-raised-cosine filter, Digital filter, Prototype filter, Adaptive filter