Texture Segmentation by Grouping Ellipse Ensembles via Active Contours

Texture plays an important role in human visual perception and offers crucial cues for solving a wide range of computer vision problems, such as image segmentation or scene analysis. The segmentation of texture is a key problem in computer vision and image understanding, the objective of which is to partition an image into several regions characterized by homogeneous texture attributes. Over the course of the past 40 years, numerous studies have been performed for texture segmentation, see [2, 3, 9, 10, 13, 14]. In this paper, we address the issue of structured texture segmentation, starting with the assumption that textures are statistical ensembles of local image structures, also known as textons [9, 16]. The study presented in this paper is inspired by works in mathematic morphology, more precisely, by granulometry [11, 12], which characterize textures relying on responses to morphological filtering with userspecified structuring elements of increasing size. The segmentation of synthetic and simple textural images can be achieved by partitioning the image according to some statistics of the granulometry [4, 5, 11], but it fails at describing complicated and highly structured textures [5]. Instead of using structuring elements and improving the discriminative powerful, alternative approaches have been proposed to analyze textures based on connected operators which perform directly on the level lines of images, see [6, 7, 15]. The main motivation of this paper is to investigate the granulometry-like approach in the context of texture segmentation. Along the line of textons, we first suggest to represent textures by a tree of ellipses, which are derived from the level lines of images and can be regarded as explicit textons. As we shall see, the tree of ellipses of an image can be computed rapidly and efficiently, thus the proposed approach can overcome difficulties in the detection of texture primitives or texture elements as encontered in [1, 13]. Based on this representation, textures are subsequently characterized by geometric properties of and by relationships between these ellipses. Thus, the modeling of a texture u is reduced to the modeling of the tree of ellipses (E ,T ), as