Research On Image Co-segmentation Method Based On Target Diversity

By obtaining the prior knowledge of the foreground similarity among images,the co-segmentation method can achieve target segmentation well for multiple images which have common targets.It makes up for the problem that the traditional single image segmentation method shows the lack of target segmentation performance in the absence of artificial interaction.Therefore,it becomes a hot topic in the field of image segmentation.However,the existing co-segmentation method is sensitive to the target which has different shape or color.How to eliminate the influence of the diversity of the target is a difficult problem to be solved in this field of co-segmentation.In this paper,we studied co-segmentation methods which are applied to different target scenes.As co-segmentation method for multiple common targets lacks of integrit y and precision,a co-segmentation method based on the maximum common subgraph and GrabCut is proposed by eliminating the redundant background area and growing the segmentation area of pixel-level image segmentation.In addition,to solve the problem of node redundanc y for the directed graph co-segmentation method which can be used for the segmentation of the common targets which have different colors,a directed graph co-segmentation method based on improved node generation rules is proposed.The specific work of this paper is as follows:In this paper,a co-segmentation method based on maximum common subgraph and GrabCut is proposed.First,the maximum common subgraph matching is carried out on an image pair to obtain the initial segmentation results.Next,we eliminate small background areas.Then,the segmentation results are growed by obtaining the outer rectangle region of th e foreground region and combining the single image background evaluation method to obtain more complete results.Finally,the GrabCut method is used to segment the previous results on pixel level to achieve more accurate segmentation.The accurac y performed on the iamge pair dataset reaches 87.75% and the recall rate is 87.71%.It proves that the proposed method is effective and feasible.A co-segmentation method based on improved node generation rules is also proposed in this paper.We eliminate redundant nodes by combining threshold method and single image background evaluation method to reduce the time complexity of the algorithm.In the process of constructing the directed graph,we calculate the edge weight by combining the color features,contour features and image salienc y information of the images to solve similar background problems.We tested our method on MSRC and iCoseg dataset.The experimental results show that it can perform segmentation well and the average node size of the graph is reduced by 40.37 and 34.83% respectively,the time cost for solving directed graph is reduced b y 38.52% and 32.92% respectively.