Research On Collaborative Optimization Of Pixel-level And Object-level Markov Random Field Models

The Markov Random Field(MRF)model is a probabilistic graphical model,widely used in image segmentation.The MRF model was originally used for pixel-level image segmenta-tion and later extended to the object level.However,pixel-level MRF and object-level MRF have their own advantages and disadvantages:The pixel-based MRF model can conveniently use regular spatial context information.But the calculation of the pixel-based MRF model is time-consuming,and its small neighborhood system also limits the segmentation accuracy;The object-based MRF model is based on over-segmented objects,which can reduce the number of nodes in the graph model,obtain more macro and complex spatial information in larger neigh-borhood systems,and reduce the computation time.But the effect of the initial segmentation region is large,the spatial relationship between regions is irregular,and it is difficult to define the neighborhood system.In order to complement each other's advantages of pixel-level and object-level MRF mod-els,this paper starts from different perspectives,combines the two models,and proposes two new models to achieve the purpose of optimizing the MRF random field model.Its main work is as follows:1.Starting from the labeled data and drawing on the idea of collaborative training,this paper proposes an MRF model(UMRF-LD)that combines pixel level and object level based on label data.First,the pixel-level MRF model(ICM)is used to obtain the optimal segmentation result at the pixel level as the initial segmentation.This initial segmentation is assigned to the initial segmentation of the object-level MRF model(OMRF)to obtain the optimal segmentation results at the object level;Then use one of the method's segmentation results as the initial value of another method to alternate iterations;The two methods are gradually optimized and progress together to obtain the final segmentation result.2.Starting from the observed data,combine the pixel features with the regional features through a likelihood function,this paper proposes an MRF model(UMRF-OD)model combin-ing pixel level and object level based on observation data.Firstly,the pixel features and region features of the observation data are extracted,and the region features are defined by taking into account the macro information such as the size and shape of the over-segmented objects when extracting the region features;Secondly,the likelihood function is decomposed into the product of the pixel likelihood function and the object likelihood function.A new probabilistic rea-soning principle is proposed to integrate the different types of likelihood information and space constraint relationships;Finally,the segmentation result is optimized by iteratively updating the posterior probability of the proposed model.In order to evaluate the effectiveness of these two models,we performed tests on different datasets and some synthetic texture images.We compared these two models with other experi-mental results based on the MRF model.The data shows that the model proposed in this paper is more effective in image segmentation.