Research And Application Of Image Semantic Segmentation Algorithm Based On Conditional Random Field

Image Semantic Segmentation is to divide an image into groups of pixel regions with certain semantic meanings,and identify the categories of each region to obtain an image with pixel semantic tags.With the rapid development of Internet technology,image semantic segmentation has gradually matured,and its research results are widely used in automatic driving system,robot navigation and intelligent medical treatment.In recent years,Conditional Random Field(CRF)based on probabilistic graph model has been widely studied and applied in image semantics segmentation because of its ability to fuse multiple features and express context information in a unified probability framework.CRF model is a discriminant probabilistic undirected graph model,which is used to mark sequence data and has strong probabilistic reasoning ability.It can effectively deal with the relationship between multiple features without the assumption of independence,and can deal with the long-distance trigger relationship in the observation sequence,which other image segmentation models can not do.Because the existing Pairwise CRF image semantics segmentation model has limited expressive ability,it is difficult to obtain the desired semantics segmentation results.Compared with the CRF model which only considers the local neighborhood,the fully connected CRF(FCCRF)model further considers the relationship between a single pixel and all other pixels,establishes the dependency relationship between all the pixels in the image,and introduces the spatial global information of the image using the interdependence relationship between the pixels,so as to obtain better semantics segmentation results.Therefore,this paper studies image semantics segmentation based on FCCRF model,and the main research contents include the following aspects:(1)Aiming at the ambiguity of edge in full-connected CRF model for image semantics segmentation,a FCCRF image semantics segmentation model based on global probabilistic boundary constraint(gPbEC-FCCRF)is proposed.Firstly,the texton Boost classifier is used to acquire the basic structural features of the image and establish the one-dimensional potential energy terms of the FCCRF model.Secondly,the edge contour features of the image are extracted by the gPb edge detection algorithm and fused into the Gauss edge potential energy of the FCCRF model to establish a new point-to-point potential term of the image segmentation model.Finally,under the framework of the FCCRF model,a new edge potential term is established.FCCRF image semantics segmentation model of energy items.The experimental results show that the proposed model can better preserve image edge information and have better semantics segmentation results.(2)Aiming at the problem that FCCRF model difficult to accurately describe the high-dimensional complex features of natural images,a high-order FCCRF image semantics segmentation model based on robust P~n Potts model(RHOFCCRF)is proposed.Firstly,high-order potential energy can be used to introduce more complex image features.Based on the FCCRF model,a robust P~n Potts model is used to establish high-order potential energy terms of the image segmentation model.Secondly,a robust high-order FCCRF image semantics segmentation model is established by weighting the high-order energy terms with the one-dimensional potential terms and the point-to-point potential terms of the FCCRF model.Finally,experiments show that the proposed model can accurately model high-dimensional complex features of complex natural images and obtain more accurate semantic segmentation results.(3)In order to better maintain the low-dimensional edge information of the image and to describe the high-dimensional complex features of the natural image more accurately,the gPb edge energy constraint is combined with the high-order potential energy function,and the image semantic segmentation based on gPbEC-RHOFCCRF is proposed.The model is applied to the road scene semantic image segmentation task.The experimental results show that the proposed model can effectively improve the accuracy of road scene image semantic segmentation.