Median and neural networks hybrid filters

• Published in: Proceedings of ICNN'95 - International Conference on Neural Networks Vol. 1; pp. 580 - 583 vol.1
• Main Authors: Taguchi, A., Muneyasu, M., Hinamoto, T.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 1995
• Subjects: Backpropagation algorithms; Gaussian noise; Image processing; Neural networks; Nonlinear filters; Prototypes; Signal processing; Statistics; Working environment noise
• ISBN: 9780780327689; 0780327683
• DOI: 10.1109/ICNN.1995.488740

In this paper we present a new nonlinear filter with neural networks for signal processing in a mixed noise environment, where both Gaussian noise and impulsive noise may be present. Mean filters can effectively remove the Gaussian noise and order statistics filters can effectively remove the impulsive noise. However it is difficult to combine these filters to remove the mixed noise in an image processing environment without blurring the image details or edges. In order to remove a mixed noise while preserving edges and details, we develop a novel prototype filter which is composed of two stages. The purpose of the first and second stages is to remove the impulse noise and Gaussian noise, respectively. The prototype filter can be shown by a network structure. This network can be extended and generalized to the median and neural networks hybrid (MNNH) filter. The coefficients of the MNNH filter are learned by the backpropagation algorithm.