Research Of Image Segmentation Algorithm And Its Applications Based On Dense Correspondence

Image segmentation is the core problem in computer vision,which extracts the object of interest by dividing an image into disjoint regions according to the different regions showed obvious differences.It is widely used in industrial,medical,military and other fields.There is no universally applicable optimal method for segmentation because of the complexity of image characters,such as brightness,color,texture and so on.Therefore,we propose a novel image segmentation algorithm based on dense correspondence,which is to simplify and change the image appearance through the fast calculation result of correspondence pixels—pixel flow field,solving the difficulties of extracting target in complex background and the bottleneck of segmentation efficiency.This paper discusses the related algorithm—imag segmentation and dense correspondence,comparing and analyzing the application scenarios and limitations of the existing algorithms.Then we introduce the proposed Grab Cut-Like,which combines the dense correspondence flow field and interactive Grab Cut,segmenting the foreground object from the image background quickly and accurately.The main research contents are as follows:1.Briefly described the basic theory of image segmentation,introducing and analyzing several common segmentation methods.2.Illustrate some dense correspondence algorithms so as to enhance the appearance consistency of the correspondence pixels and the geometric smoothness of the adjacent pixel.3.Propose an image segmentation algorithm based on dense correspondence,establishing the dense correspondence of the input image and its mask by the calculated results—pixel flow field and warping the mask to the original image which produces a target mask.Then conducting the subsequent segmentation tasks based on Gauss Mixture Model.4.Improve the single warp model in original dense correrpondence,warping more than one mask and normalizing to generate the most close to the Ground-Truth,which effectively improves the accuracy of segmentation.5.Propose the Grab Cut-Like segmentation algorithm with the improvement of Grab Cut interactive segmentation.We use the mask image to initialize the Grab Cut without manual interaction.According to the foreground/background region provided by mask image,realizing the automatic selection and segmentation,which enhances the accuracy and efficiency of segmentation effectively.Experiments were conducted on the two datasets: Caltech-101 and MSRC.We verify the robustness of our segmentation algorithm by comparing and analyzing the experimental results.Also illustrated the Grab Cut-Like could solve the multi-target image and the challenging segmentation tasks in the complex background,effectively expanding the application areas.