A neural network for texture classification

This paper describes a new method, based on recovery of directionality from compositional textures. The technique uses features obtained as projection coefficients of the flow field estimated from texture onto suitable elementary basis vector fields. A neural network is used for both the estimation of the best projection into the feature space and the segmentation of the original image into homogeneous regions. The technique has been verified on both natural and artificial textures; some results obtained on leather for defects detection and classification in manufacturing industry are shown.