Adaptive fuzzy morphological filtering of impulse noise in images (Q1590495)

The aim of this paper is twofold: to introduce the implementation of fuzzy morphological operators using neural networks, and to propose a training algorithm for the optimization of Structuring Elements (SEs) to an image. A procedure for designing fuzzy morphological filters for the removal of impulse noise from binary and multi-level images is given. Convergence of the optimization for specific SEs, and some general convergence results are also discussed. In order to avoid convergence to local minima, it is shown that a flat structuring element should be used as the initial one. The optimal SEs obtained from the supervised or unsupervised training methods provide a structural characterization of the images, proved to be useful in many applications, including noise removal. The performance of the fuzzy morphological filters in removing impulse noise in multi-level and binary images is illustrated and compared to other existing procedures.