Image Object Segmentation Based On Hierarchical Conditional Random Fields

Image object segmentation is an important research topic of computer vision, as well as a critical step inunderstanding and analysis of images. So far, there are many kinds of methods, the mains methods useConditional Random Fields framework. These methods are performed in units of pixels, including featureextraction and training model, which exists a large number of redundant computation. Besides, how to usesemantic information of the image is a research focus point. Semantic information can greatly improve theaccuracy of the image segmentation.For the above problems, a novel approach of establishing hierarchical CRFs (Conditional Random Fields)for image object segmentation has been proposed in this paper. This method treats a super-pixel as a processingunit, and uses the super-pixel-based textonboost features. Besides, this paper presents an image segmentationmethod based on Bayesian harmony degree, the method can automatically select the appropriate number of region.The obtained area contains the same object as possible, which can reflect certain semantic information. Wecombine this region to define the higher order potential. This potential can build a level constraint relationshipbetween super-pixel and region, and then create a hierarchical CRFs model. In this paper, the hierarchical CRFsparameters are trained by using the features based on super-pixel and region. Finally, we obtain the final imageobject segmentation results by minimizing the energy of the established hierarchical CRFs.Finally, we evaluate the performance of the proposed method on the MSRC dataset. Our method iscompared with the image object segmentation based on super-pixel and pixel respectively. The Experiment isanalyzed from two aspects of model training time and the accuracy of image object segmentation. The results onthe benchmark dataset show that the proposed method achieved better results in both the rate of training modeland the accuracy of image object segmentation.