Unsupervised classification of texture images by gray-level spatial dependence matrices and genetic algorithms

Recognition of objects and regions of interest in digital image processing often relies on texture classification. The source image is divided according to a rectangular grid to form textured regions each of which is characterized by some numerical significant measure called feature. A new approach is introduced that uses the gray-level spatial dependence matrices and the genetic clustering with unknown K algorithms to locate sets of homogeneous regions and enhance the discrimination amongst them. There is no need to select and compute complicated features transforms as the procedure is based on the optimal weighting of the simple basic features. A simulation experiment is performed using the well-known Brodatz textures to demonstrate that the new procedure is able to define well separated clusters according to the principle of strong internal cohesion and high inter-clusters separation.