Image Segmentation Research Based On Level Set Method

Image segmentation is a kind of technique that partitions an image into uniform and non-overlapping regions based on some similar measure and extracts those interesting objects. Level set method has become an important direction in the field of image segmentation, and has manifested the favorable performance. However, level set segmentation method is still developing now, and its theories and applications need further research. A fatal shortcoming of existing level set methods is that they usually request human given initial contour and sometimes the initialization effect is also not good enough. Furthermore, they also behave with poor generality. Therefore, in consideration of some drawbacks of traditional level set segmentation methods, such as the human intervention demand, slow segmentation speed and defective segmentation result and so on. This thesis pays more attention to the automatic preprocessing of the initial contour, and puts forward a new level set segmentation algorithm. The main works in this thesis can be summarized as follows.1. Traditional level set method generally requires manual given initial contour, which is randomness and uncertainty. It potentially leads to undesired segmentation effect, more evolution iterations, and longer running time and so on. In order to improve the segmentation effects, this thesis puts forward two autonomous initialization methods and makes the initial contour be close to the edges of those targets. It will provide a good foundation for level set evolution.(1) In this paper, an autonomous gradient-based approach is proposed for deciding the initial contour of the level set method, by amplifying the difference of gradient value between edge and non-edge. The data of edge and non-edge perform obvious difference, and then search a threshold which can be used to partition the edge and non-edge. The threshold meets keeping the minimum sum of the variance of two parts. The peripheral contour of those points smaller than the threshold is considered as the initial contour, which can be directly evolved without artificial positioning. The ideal results and the short performed time shown here demonstrate the potential of our innovative approach.(2) We propose a global-based approach for deciding the initial contour of level set method. The intensity difference data for entire image are obtained by the mean intensity of image, which is regarded as the threshold. The intensity difference for some point indicates which object it belongs to. However, a lot of images are usually noisy or blurry, and those points close to zero may belong to the wrong classification. In this paper, we set two special parameters to adjust some deviant points, and the midline of positive and negative is taken for the initial contour. The experiments confirm that the autonomous approach for deciding the initial contour can greatly improve the result of level set segmentation.2. In this paper, we propose a new global-based level set segmentation approach. We consider the statistical information of three different regions to construct the level set model, including contour, inside contour and outside contour. The evolution function is constructed by the statistical information of the three parts. Compared with traditional global-based level set methods, such as the CV model, the proposed method improves segmentation effects of blurry images due to designing statistical information of the zero level set region. Then, we combine the autonomous initialization approach and the proposed level set segmentation method. The combination of the two methods for different stages makes us to get more desired segmentation effect.In this paper, we extend and improve the level set method for image segmentation. A great deal of experiment has demonstrated the validity and feasibility of the proposed methods. In the future, we will study further in the fields of multiphase image segmentation, texture or color image segmentation, visual target tracking and so on.