Learning type of median and mean hybrid filters and a synthesis of its learning signal.

Aiming at the elimination of Gaussian noise and impulsive noise, the authors have proposed the median filter and mean hybrid (MMH) filter, which combines the mean filter, which is useful in eliminating Gaussian noise, with the median filter, which is useful in eliminating impulsive noise. This paper proposes the learning MMH (LMMH) filter with the learning function, where the mean filter and the median filter in the MMH filter are replaced by the FIR filter and the order statistics (OS) filter, respectively, which are more general filters. The LMMH filter is optimized by learning, and the information concerning the position and the order in the training signal are well reflected on the LMMH filter, which improves the accuracy of the signal restoration, compared to the MMH filter. This paper further proposes a method by which the learning signal to be used in the learning of the LMMH filter can be constructed based only on the signal to be processed, which is degraded by the noise. By applying both the LMMH filter and the method of constructing the learning signal, a precise method for learning of the LMMH filter is provided, even in actual situations where the original signal for the signal to be processed is unknown. It is shown through an application example that the LMMH filter obtained by the learning exhibits a stable and better performance than the MMH filter. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 83(3): 95–107, 2000 [ABSTRACT FROM AUTHOR]