Texture analysis using an adaptive neuro-fuzzy network and fractals

Feature extraction and pattern classification are two major steps in image data processing. For feature representation a fractal geometry based vector, the components of which are multiresolution fractal dimensions based on the fractional Brownian motion models, is proposed. The vector can be calculated using two dimensional covariance function of the textured image under consideration. For pattern classification an adaptive neuro-fuzzy network containing blocks, each block being a fuzzy network with center average defuzzifier, product inference rule, singleton fuzzifier, and Gaussian membership functions. The blocks can be trained separately. It was proven, that the neuro-fuzzy classifier has better performance than a neuro classifier because it is more realistic and natural to use fuzzy methods than the hard classification techniques, that is due to the possible presence of partial characteristics of multiple texture classes present in the training vectors.