The Study Of Image Segmentation Algorithm Based On Regional Level Set Methods

Image segmentation is one of the most crucial and challenging tasks in computer vision and pattern recognition, which is very useful for object detection, object recognition, image retrieval, scene analysis, medical image processing, and video surveillance. Up to now, in this field, a wide variety of methods have been proposed including the region merging based methods, the graph based methods, and the active contour model (ACM) based methods. Among them, the most successful and important one is the level set method, which uses a deformable curve front to detect the boundary of an object and is based on desirable mathmetic theory. The region-based models have attracted more attentions due to the ability of capturing region intensity information. Based on regional level set methods, this paper makes lots of researchments for the problem of complicated image segmentation, including images with severe intensity inhomogeneity, nature texture images, and so on. Most important of all, some effective methods are proposed to improve the segmentation results of complicated images.1). This paper presents a novel level set method based on piecewise constant approximation and multi-scale structure. Rather than detecting accurate intensity inhomogeneity, we reformulate the intensity inhomogeneity model implicitly via approximation of piecewise constant image so that piecewise constant segmentation criterion becomes applicable. Further, considering the variance of local intensity distribution and reliability of local regions, we propose a Gaussian pyramid convolution strategy to construct local neighborhoods with different scales. So, the multi-scale structure is obtained and used to model the piecewise constant image. And then, according to piecewise constant segmentation criterion, we construct the constant descriptors to describe approximated piecewise constant image under multi-scale structure so that the multi-scale structure data term is obtained. To blend the derived information from multi-scale structure, we incorporate the multi-scale structure by computing mean of data terms under different local neighborhood scales. Finally, we utilize level set method to construct the energy functional based on multi-scale structure. The experimental results demonstrate that the proposed method yields results comparative to and even better than the classical models for segmenting images with intensity inhomogeneity.2). A novel level set segmentation model integrating the intensity and texture terms is proposed to segment complicated two-phase nature images. Firstly, an intensity term based on the global division algorithm is proposed, which can better capture intensity information of image than the Chan-Vese model (CV). Particularly, the CV model is a special case of the proposed intensity term under a certain condition. Secondly, a texture term based on the adaptive scale local variation degree (ASLVD) algorithm is proposed. The ASLVD algorithm adaptively incorporates the amplitude and frequency components of local intensity variation, thus, it can extract the non-stationary texture feature accurately. Finally, the intensity term and the texture term are jointly incorporated into level set and used to construct effective image segmentation model named as the Intensity-Texture model. Since the intensity term and the texture term are complementary for segmentation, the Intensity-Texture model has strong ability to accurately segment those complicated two-phase nature images. Experimental results demonstrate the effectiveness of the proposed Intensity-Texture model.3). A local maximum description difference (LMDD) feature is proposed to determine the maximum response of multi-scale high-pass filter for each image pixel and incorporated into level set energy functional to construct the multi-scale local models. Firstly, the multi-scale local region descriptors are constructed to describe the local regions with different scales. Then, the LMDD feature is extracted based on the multi-scale local region descriptor associated with high-pass filtering. Secondly, the LMDD feature is incorporated into three typical level set models with Chan-Vese (CV)-like structure. Here, the LMDD feature is utilized to weaken the influence of the intensity inhomogeneity and enhance the contrast between object boundary and background regions. Finally, the multi-scale segmentation is performed by minimizing the new formed level set energy functional. The optimal local region scale for each pixel is determined by LMDD feature and may be changed with the contour evolution, which can greatly enhance the ability of evolving contour to approach the true object boundary. The experimental results demonstrate the effectiveness and efficiency of the proposed multi-scale local models for segmenting images with severe intensity inhomogeneity.4). We propose a novel level set method for image segmentation which introduces the moment competition and semi-supervised clustering ideas into the energy functional construction. Different from the traditional force competition based level set methods, the moment competition is adopted to drive the contour evolution instead of force competition. Firstly, a so-called Three-Point labeling scheme is proposed to manually label three independent points (seed pixel set) on image. Similar to sample labeling in semi-supervised clustering methods, the seed pixel set is used to construct the clustering distance for each image pixel. Then, the force arm can be computed based on the combination of intensity clustering distances. Unlike the traditional method, the force is generated from the global intensity difference and weighted by force arm. Finally, the moment is constructed and incorporated into the level set energy functional to drive the evolving contour to approach the object boundary. In our method, the force arm can take full advantage of the Three-Point labeling scheme to constrain the moment competition. Besides, the global statistical information and the sample labeling information are successfully integrated which makes the proposed method more robust than traditional force competition methods for initial contour placement and parameter setting. Experimental results also show the priority of the proposed method on segmenting different types of complicated images such as noisy images, three-phase images, images with intensity inhomogeneity and texture images.