A Self-Building Fuzzy Neural Network for Noise Reduction in Images

摘要

Prior research indicates that, by utilizing fuzzy sets, it is possible to create fuzzy 3×3 or 5×5 block mappings to perform various types of low-level image processing. In the case of noise reduction, blocks of an original noisy image are mapped to blocks in the processed ("clean") image. Furthermore, it then became feasible to train a radial basis function network to perform image processing. One drawback of this approach is that the mappings must be hand-made, meaning that the some sets of mappings may be incomplete or redundant. In this paper we present a radial basis function network that determines its own mappings by systematically building its hidden layer as it trains. The resulting structure is able to detect and reduce noise in images with approximately 90% accuracy.