Unsupervised Image Segmentation based on the Multi-resolution Integration of Adaptive Local Texture Descriptors

The major aim of this paper consists of a comprehensive quantitative evaluation of adaptive texture descriptors when integrated into an unsupervised image segmentation framework. The techniques involved in this evaluation are: the standard and rotation invariant Local Binary Pattern (LBP) operators, multichannel texture decomposition based on Gabor filters and a recently proposed technique that analyses the distribution of dominant image orientations at both micro and macro levels. These selected descriptors share two essential properties: (a) they evaluate the texture information at micro-level in small neighborhoods, while (b) the distributions of the local features calculated from texture units describe the texture at macrolevel. This adaptive scenario facilitates the integration of the texture descriptors into an unsupervised clustering based segmentation scheme that embeds a multi-resolution approach. The conducted experiments evaluate the performance of these techniques and also analyze the influence of important parameters (such as scale, frequency and orientation) upon the segmentation results.