Texture classification using neural networks and discrete wavelet transform

Describes a method for classifying textured images using neural networks and discrete wavelet transform (DWT). In this method, a multiresolution analysis is applied to textured images to extract a set of intelligible features. These extracted features, in the form of DWT coefficient matrices, are used as inputs to four different multilayer perceptron (MLP) neural networks and classified. Generalization performance is improved when a locally connected, weight-sharing network topology is utilized, thus drastically decreasing the number of free parameters during training. This architecture takes advantage of the quasi-periodic nature of the textured images. A novel voting network scheme is also employed to achieve a system classification result from the four networks. The efficacy of the algorithm is demonstrated using real-world textured images