The Research Of Image Segmentation Based On Local And Global Constraints

With the coming of the digital age,Digital images become the main carrier of a large amount of information,how to extract the information from the image becomes a hot issue.In computer vision,image segmentation is an important means to get people's interested information from the image,which is the key link of image analysis and understanding.Most of the traditional image segmentation methods use only the local consistency of the image apparent feature,which is suitable for scenarios where the foreground and background characteristics are easy to distinguish.However,when the regional boundary of object is blurred and there is occlusion and the background is complex,these methods have some limitations.Aiming at this problem,this paper introduces the object shape as a global constraint,and proposes a joint segmentation model combining CRF with ShapeBM model.CRF constructs local constraints through the second order potential(marginal energy),ShapeBM represents the image global constraints by modeling the shape of the object.The experiments on the Penn-Fudan Pedestrians dataset,Caltech-UCSD Birds 200 dataset and Labeled Faces in the Wild(LFW)dataset demonstrate that the joint model is more effective and efficient than others.The main contents of this paper are as follows:(1)The boundary of superpixel usually is the real object boundary,which can improve the segmentation accuracy.Superpixel are used to replace pixels,which can reduce the complexity of the calculation.This paper proposed CRF segmentation model based on superpixel which is used as a segmentation model based on local constraints.(2)The CRF model use only image local constraints,there are some limitations in the segmentation of complex scenes.The ShapeBM modeling object shape is introduced into the CRF model as a global constraint,which effectively overcomes its limitations and improves the application of the model.On this basis,this paper proposes a pooling technique to solve problem that the number of nodes the superpixel layer of the CRF does not correspond to that of the input layers of the ShapeBM.(3)During the training process in the model,this paper gives the relevant training algorithm,which uses the pre-training method to obtain the initial model parameters better than the random initial parameters,so that the speed of the model training convergence is accelerated.In the process of inference,the joint model is based on the combination of CRF and ShapeBM,so the infer superpixel label is not only related to CRF,but also related to the hidden node of ShapeBM.This paper presents the CRF model and ShapeBM hidden node joint inference algorithm to infer the superpixel label.