Image Cosegmentation Method Based On Minimum Fuzzy Divergence

Image segmentation is a research problem in the field of computer vision,and a great number of research methods have been proposed in recent years.With the fast development of the Internet leading to the rapid growth of image data,segmenting foreground objects from images has been an active topic which is a low-level step in many practical applications such as intelligent systems.Researchers are no longer satisfied with single image segmentation,and gradually pay attention to a weakly supervised segmentation method named cosegmentation.In the cosegmentation method,the input image sets share the same or similar foreground objects.However,the complex background of the image and the changes of the foreground increase the difficulty of segmentation.Therefore,designing a flexible and accurate foreground segmentation algorithm has become the key point and difficulty of research.The research is focused on designing a new foreground similarity measurement method,which gives enough consideration to the foreground area information.And the foreground similarity measurement method is integrated into the energy functional to improve the region-based active contour model.Finally,the experiment results are compared with the traditional cosegmentation methods to prove the effectiveness of the improved method in this paper,which include qualitative results and quantitative results.The region-based active contour model is widely used in cosegmentation problems.Its principle is mainly to integrate the inherent nature of the curve and image area information into an energy function which is updated through the level set function.Finally,when the energy function is minimized,the curve evolves to the best object boundary.Based on the traditional active contour model,this paper solves the problem of cosegmentation of image pairs sharing the same object.The main content and innovations of the thesis are described as follows:(1)In the area-based active contour model,the foreground similarity measure is the key technology of the entire model as the global term of the energy functional.In our model,the color histogram method is first used to simply classify the pixels of the input image to obtain a rough and inaccurate initial contour.But the color histogram has a limitation that it is unable to determine the spatial position of each color.Therefore,the fuzzy divergence is proposed as a new measure of foreground similarity.Fuzzy divergence has the property of measuring the similarity between two sets so our model incorporates it and the energy function.Considering the foreground similarity and background consistency between multiple images,the pixel information covered by one image is integrated into the energy function of another image to promote the dynamic evolution of the curve and enhance the robustness to the initial contour.According to the minimum fuzzy divergence criterion,the minimization of the energy function means that the curve evolves to the ideal contour boundary.Experiment results show that the cosegmentation model proposed in this paper can achieve better segmentation results than traditional methods,and the error rate is lower than that of traditional cosegmentation methods,and the robustness is stronger.Among them,the maximum IOU value is 0.98495,the maximum Precision value is0.97479,the maximum Dice value is 0.9911.(2)In many practical applications,the object sets researched always have fuzzy properties that cannot be accurately classified,such as image segmentation,image denoising.Fuzzy set theory is precisely used to describe this phenomenon with uncertainty.Nowadays,more and more fuzzy set related methods and theories are applied to the problem of image segmentation.In the past,fuzzy divergence was often used for single grayscale image segmentation.The paper puts forward the point for the first time that fuzzy divergence is used to deal with color image cosegmentation.In fuzzy set theory,membership functions are used to describe fuzzy sets.Researchers have successively proposed a variety of membership functions,but the selection of membership functions still has a certain degree of subjectivity and needs to be specifically analyzed according to the distribution of pixels.In this paper,the membership function based on weighted average operator which combine the Gamma-type membership function with the logarithmic membership function is integrated into the exponential fuzzy divergence.The experiment results show that the proposed membership function achieves a better segmentation effect.Among them,the maximum IOU value is 0.974,the maximum Precision value is 0.981,the maximum Dice value is 0.991.This paper selects the cluster-based cosegmentation algorithm and the reward strategy-based cosegmentation algorithm to compare with our model.Qualitative and quantitative experiments are performed on the classic i-Coseg database,MSRC database and the Caltech-256 Object Categories database.And three evaluation methods including IOU,Precision and Dice are used to evaluate the segmentation results.Experiment results show that the model in this paper achieves a better cosegmentation effect for image pairs with the common object,with lower error rate and stronger robustness.