A Novel Structure Tensor Modulated Chan-Vese Model for Texture Image Segmentation

In this paper, a modified Chan-Vese model is proposed which is suitable for texture segmentation. The framework is based on a tensor-processed linear structure tensor (LST) as the input to the variational model. A LST is a function of the image gradient which highlights the dominant gradient direction over a neighborhood, while providing a degree of coherence. In the proposed model, the original image is considered as an additional information channel in computing the structure tensor. This nullifies the dislocation of edges to some extent, which otherwise occurs when the structure tensor is processed by the Gaussian smoothing operator. It also enhances the vector magnitude to an appropriate order without losing the directional property due to dislocation. Additionally, an integrated stopping condition is proposed which enables the algorithm to decide the number of iterations adaptively and thus achieve automatic segmentation. The proposed model is applied on suitable test images from the Berkeley Image Database and the results are compared with conventional Chan-Vese, Local Binary Fitting, Gabor-based Chan-Vese and LST integrated models. Quantitative and qualitative analyses are presented to illustrate the potential of the proposed model.