Texture Analysis Using Neural Networks and Mode Filters

This papei describes experiments in which the Laws' approach to texture analysis is augmented by artificial neural network classifiers followed by mode filters. Analysis of image classification data confirmed that averaging filters should be carefully chosen to obtain the best balance between remanent texture noise and blurring of class boundaries, smaller mask sizes being required for segmentation vis-a-vis classification. Preliminary image normalisation was found to be generally advisable, unless the training data is fully representative of the test data, while mode filters are valuable for consolidating the output images, e.g. by eliminating patches of misclassification. Indeed, mode filters seem to constitute an under-exploited technique in image analysis as a whole.