The effect of features reduction on different texture classifiers

Texture analysis continues to evolve as a feature measurement technique to analyze and classify image data. Texture features could be derived using various approaches. One of the most important problems in the field of pattern analysis and texture classification is the problem of finding an appropriate set of features-vector that represent observations with reduced dimensionality without sacrificing discrimination power. This work aims to examines the role of principal components analysis (using PCA neural network) in achieving feature reduction in texture analysis, and attempt to draw a conclusion about the effect of feature reduction on the classification accuracy of natural-textured images (from Brodatz album) and if the feature reduction has the same effect on the behavior of different classifiers. To do this, different sets of features (statistical, statistical with radial/angular, spectral), before and after reduction, are used to train ACON based (All- Classes-One-Net) neural networks (Kohonen , LVQ, BP, PCA, and ART2) for texture-classification, then the classification accuracy of the NN with reduced and non reduced feature sets are compared and assessed.