Unsupervised Texture Image Segmentation Based on Local Features

Abstract

The article is devoted to the problem of unsupervised segmentation of texture images. Texture segmentation can be used in the analysis of satellite, geological, medical, biological images. Texture segmentation algorithms have many advantages: lightness, speed, the possibility of unsupervised implementation, simpleness of automatic labeling of data sets, a set of target tasks. Model with the following steps was chosen to solution. After a simple pre-processing of the raw image, the local vector texture characteristics of the each point of the image are calculated. For this goal used a bank of convolutional Gabor filters and empirical curvelet filters that produce frequency splitting. Computed local histogram aggregation. Then, clustering is performed based on the construction of a feature matrix and singular decomposition. All levels of segmentation automation up to fully automatic have been explored. All algorithms are implemented and tested on prepared synthetic and real data.