Texture segmentation using joint time frequency representation and unsupervised classifier

Proposes a new texture segmentation method based on joint time frequency representation and an unsupervised neural network classifier. The proposed method uses a filter bank with variable parameters to sample the frequency plane. Since the outputs of these filters are believed to reflect the frequency distribution of pixels, they can be treated as texture features. A multiresolution frequency sampling technique is developed to help determine the filter bank parameters, so that the extracted features are optimal under a certain information cost function. A self-organizing neural network is also adopted to implement unsupervised classification of pixels according to their texture features. Some experimental results show that the presented method is efficient and robust especially in the case of natural texture.