Research On Texture Segmentation Via Sparse Representation With Geometric Constraints

Texture segmentation is a classic problem in computer vision,which is to divide a texture-dominant image into different texture regions.As a key technology,texture segmentation has been widely used in many visual applications,such as automatic robot navigation,remote sensing image analysis,and medical image analysis.One key in texture segmentation is to correctly find out the texture elements for different texture regions.Such a task can be viewed as the process of discovering the low-dimensional structures of high-dimensional texture images,which can be efficiently done by the sparse representation with dictionary learning.In this paper,we focus on developing sparse representation models for texture segmentation.In real scenarios,the expected label images outputted by texture segmentation often show some geometric properties,such as piecewise constancy of labels and continuity of region boundaries.However,traditional sparse representation models cannot characterize such geometric properties.The main content of this paper is to propose effective sparse representation models for texture segmentation,which can well induce the geometric properties in segmentation results.Furthermore,the related texture feature extraction methods,as well as post-processing techniques are also studied in the papers.Based on the proposed models and techniques,two complete texture segmentation methods are developed,one for the case where user interaction is allowed,and the other for the case of fully automatic segmentation.The main work of this paper is summarized as follows:?1?A user-interactive texture segmentation method is proposed.The method allows a simply yet efficient way for a user to give initial labels to small texture regions.With such labels,the texture segmentation problem is transformed to a weakly-supervised patch classification problem.Then a weakly-supervised sparse representation model is proposed to exploit the given labels for segmentation.The proposed model also exploits the prior from the geometric properties of segmentation label images.The piecewise constancy prior of labels is introduced by controlling the7)₀ norm of wavelet coefficients of the label image.The continuity of region boundaries is enforced by constraining the label image as a fixed point of the closing operator.In order to obtain a more comprehensive texture feature as input,a texture feature extraction method based on multi-feature fusion of local histogram is also proposed.?2?An automatic texture segmentation method is proposed by extending the above weakly-supervised model to the unsupervised one via feature map quantization.To improve the performance in unsupervised setting,this paper also studies an improvement on the above feature extraction technique with a dimension reduction technique.The proposed methods were evaluated on two benchmark datasets with different criteria,and the experimental results have demonstrated the effectiveness of the proposed methods.