Research On Image Co-segmentation Method Based On Shape Weak Supervision

With the emergence of tremendous network images in recent years,the scale of digital image data presents explosive growth.Traditional single graph segmentation has been unable to meet the needs of many practical applications,co-segmentation is becoming a hot topic in computer vision.As a weak supervised segmentation method,co-segmentation could reasonably utilize the similarity of the common objects in images,and realize cosegmentation of the common object simultaneously without prior,and has a broad prospect.At present,a large number of researchers have proposed hundreds of co-segmentation methods based on different theories.However,the traditional co-segmentation methods are sensitive to the objects in shape and color,so how to eliminate the impact of target diversity on co-segmentation becomes an urgent problem to be solved.In complex scenes,meanwhile,the extracted low level image features cannot accurately describe the objects in the images,which results in inaccurate segmentation.How to improve the segmentation performance of co-segmentation in complex scenes is a challenging task.The level set method has the advantage of flexibly handling the topological changes of the evolution curve and the ability of integrating the knowledge of image data information and the inherent attributes of the curves into an energy functional.Based on the active contour model of level set,this paper studies the image co-segmentation methods and proposes two improved methods as follows.(1)In order to mitigate the effect of object diversity on co-segmentation,we propose a co-segmentation level set method based on Independent Component Analysis(ICA)reconstruction.Considering the diversity of background in multiple images,this method models the common object in the image uniformly,while the background in the image is modeled separately.The internal pixel collection of the evolution curve in each image is taken as the common target,and its independent features are found through ICA.The external pixels of the evolution curve in each image were regarded as the background and their independent components were extracted by ICA.ICA reconstruction was carried out on the common target and background of the image,and the energy functional of the level set method was constructed based on the reconstruction error,and finally the cosegmentation of the image group was realized by optimizing the energy functional.The associated experiments show that our method could extract the robust features of common object.Compared with several representative co-segmentation methods,the co-segmentation based on ICA could achieve better performance on noise images,texture images,and inhomogeneous images.(2)In order to improve the performance of co-segmentation on the images with complex context,we proposes a co-segmentation method based on Data Assimilation(DA)Under the framework of LSM,this method borrows the data fusion capability from DA to introduce more constraint forms.A background field model based on Partial Differential Equation(PDE)is designed to describe the geometric characteristics of the evolution curve.An observer is designed to convert the dynamic observation information into the energy functional to constrain the evolution curve.The energy functional of the level set method is constructed,and the image co-segmentation is realized by optimizing this energy functional to update the evolution curve.The experiments show that this method could effectively improve the segmentation performance of co-segmentation in with complex context background.Compared with several representative co-segmentation methods,the method has better effect on interfered images,texture images,foreground similar images and complex background images.In brief,the two methods presented in this paper focus on the design of energy functional of the level set method,improve the existing co-segmentation method,and provide a solution for introducing external data and constraints into the level set method.In additional,the two methods proposed in this paper enrich the theory and technologies of cosegmentation and promote its practical application.