Multi-spectral Texture Segmentation Based on the Spectral Cooccurrence Matrix

Multi-spectral images are becoming more common in industrial inspection tasks where the colour is used as a quality measure. In this paper we propose a spectral cooccurrence matrix-based method to analyse multi-spectral texture images, in which every pixel contains a measured colour spectrum. We first quantise the spectral domain of the multi-spectral images using the Self-Organising Map (SOM). Next we label the spectral domain according to the quantised spectra. In the spatial domain, we represent a multi-spectral texture using the spectral cooccurrence matrix, which we calculate from the labelled image. In the experimental part of this paper, we present the results of segmenting natural multi-spectral textures. We compared the k-nearest neighbour (k-NN) classifier and the multilayer perceptron (MLP) neural network-based segmentation results of the multi-spectral and RGB colour textures.