Evolutionary procedure for the optimisation of a generic texture classifier

A novel evolutionary procedure for the optimisation of a Gabor wavelet-based multi-bank filter network for texture classification is introduced. The performance of the system is evaluated in terms of the classification accuracy. The system starts with a fine mesh of feature extractor filters which result in the highest classification rate. The system is then trimmed down by an evolutionary procedure until a preset performance criterion is reached. The performance of the system is tested using 50 natural textures from the Brodatz album. The results indicates that extremely compact feature vectors with high classification accuracy could be achieved using well-structured test data