Research Of Interactive Image Segmentation Algorithm And Its Applications

Image segmentation is a fundamental technique which plays an important role in computer vision. The segmentation goal is extracting user interested targets or objects. Nowadays, the segmentation has widely applied in many industries, such as information medical, military, manufacturers, etc. Since the complex features of images(shape, textures, brightness, etc.), the traditional segmentation techniques can’t segment the desired objects accurately with obscurity boundary or ambiguity region. Therefore, interactive image segmentation has been proposed recent years. Interactive segmentation achieves the desired object with user interactions based on high-level perception. Users empathize the interested objects by strokes or scribbles, then with the algorithms iteratively computation and process, we can obtain the final segmented object. Most of the present interactive segmentation algorithms are proposed based on the classic graphcut algorithm, the others are mainly region-based or edge-based algorithms.This thesis reviews the basic knowledge about interactive segmentation, studies the current popular interactive segmentation algorithms, analyzes the corresponding advantages and disadvantages of each algorithm. Then, we combine the generative model of multi-label with the multi-layer graphical model, propose a new modified nonparametric high-order algorithm. The mainly research of this papers as follows:1. Briefly introduce the theoretical knowledge about interactive segmentation, include the importance, classification and methods.2. Analyze graph model of the traditional interactive segmentation and its related algorithms, include Graph cut and algorithms which modified based on Graph cut, at the same time, also study the concrete implantations, advantages and disadvantages as well.3. With the region and edge as the basis, we study the typical algorithms in detail, meanwhile, we also analyze two common pre-processing techniques of interactive image segmentation algorithms.4. Propose a new segmentation model which combined the Robust Pn model and multi-layer graphical model, estimate the pixels likelihood with pixel-based and region-based layer information.5. Modify the original parametric high order potential of Robust Pn model to non-parametric high order potential. We also treat the region label consistency as the soft label constraint and estimate the region likelihood with nonparametric learning method.6. With considerable experiments, compare segmentation accurate of the improved algorithm with several typical interactive segmentation algorithms, meanwhile, apply the improved algorithm to JSRT medical image dataset and MSRC-23 dataset.