Refereed full papers (journals, book chapters, international conferences)

1997

In this paper we describe an original approach to unsupervised classification of textures. Our method is based on image decomposition with a wavelet packet transform. According to our experiments, we confirm recent observations which state that a large class of natural textures can be modelled as quasi-periodic signals whose dominant frequencies are located in the middle frequency channels. We thus consider that every frequency band produces a feature image and our aim is to quantify its visual content in order to decompose only dominant channels that are able to discriminate between different textures. A vector of texture attributes is computed at every frequency band. We use for this aim statistical operators issued from cooccurrence and run length matrices. Our aim is to introduce for the multiscale analysis of textures a corresponding statistical framework. The classification results deal with good discrimination of textures. Finally, applications, results as well as future developments of our approach are discussed.