Unsupervised Texture Image Classification using Self-Organizing Maps

In this paper we report results from unsupervised texture image classification on a data set of images collected with our intelligent machine vision system for pattern recognition (assuming no a priori human vision expert knowledge is available for the image classes). The simulation of the investigated system includes four main phases: data collection and feature extraction, feature analysis, classifier training, and classifier testing and evaluation. Self-Organizing Maps (SOM) are used for classification of the collection of images into several classes, based on their features and texture characteristics. Three main experiments are conducted during this research: in the first one, all extracted features are used for training the classifiers without any statistical pre-processing of the dataset; in the second simulation, the classifiers are trained after normalization of the available data; and in the last experiment, the trained SOMs use linear transformations of the original features, received after pre-processing with principal component analysis (PCA). Each test is performed 50 times and the classification results are assessed using three commonly applied metrics, namely: accuracy rate, sensitivity and specificity. Finally, the findings of this investigation are compared with results from other authors.