Neural network based textural labeling of images in multimedia applications

In this paper the use of multilayer perceptron type neural networks in the characterization of images by texture content is investigated. The paper is focused on the effects of textural feature extraction methods on the network architecture, training performance and generalization capability when applied to indexing of images in multimedia image databases. An in depth experimental study is conducted comparing several well known textural feature extraction techniques along with a novel discrete wavelet transform based methodology. It is demonstrated that the proposed technique leads to the design and selection of multilayer perceptron architectures with the best texture classification accuracy