Image Segmentation Model Of Foreground And Background Based On The Second Order TGV With Constrains

In image segmentation,an image is usually divided into two parts,the foreground(target region)and the background(non-target region).Feature extraction and analysis of the target region obtained from image segmentation is helpful for us to have a better understanding of the nature of the target represented by the image and to provide the corresponding behavior scheme.The Total Variation(TV)image segmentation algorithm often suffer from the staircase-like artifacts and loses image details resulting in image quality degradation.Therefore,a new image segmentation method based on the Total Generalized Variation(TGV)is proposed in this thesis.In order to segment target foreground more accurately,this thesis use interactive constraint to guide the segmentation.The main work can be summarized as follows:1.As the TV model tends to produce the staircase effect,a segmentation model is proposed under the variational frame work based on the TGV.The novel segmentation model that integrates the data fidelity term of TV model and the TGV regularization makes full use of the advantage of TGV regularizer which can approximate arbitrary order polynomial function.Therefore,it can protect the important image textures and details.To further improve the quality of segmented image,an edge detection operator is introduced into the TGV regularizer to adaptively distinguish the image edge regions and smooth regions.The adaptive TGV model can reduce the diffusion to remove the impulse noise and overcome the staircase effect in the image smooth regions.2.Aiming at the shortage that complex background or the feature of foreground is similar to the background always cause segmentation errors,an constrained image foreground and background segmentation model based on the second-order TGV is presented.The interactive method is used to mark the rectangular frame of the image,so that the frame contains richer foreground and the background information outside the frame is more abundant.Mapping distance to the pixels inside the frame,outside of the frame and the edges and making the value of the distance map larger,the greater the probability of representing the foreground,the smaller the value of the distance map,the greater the probability of representing the background.According to the principle of energy minimization,this distance mapping function related to rectangular frame is introduced into the data fidelity of the TGV segmentation model.The constrained image foreground and background segmentation model based on the second order TGV not only suppressing the staircase effect and protecting the image detail features as TGV model but also has a more complete segmentation.3.The first order primal-dual algorithm to solve the model is proposed.Firstly,the feasible binary solutions of binary model are relaxed to continue functions in terms of the upper level sets function,the equivalence between minimizers of the relaxed and the binary problem is proved.Secondly,analyzing the similarity between the convex structure and the saddle point problem.Finally,a first-order primal-dual algorithm is introduced to solve the convex structure based on the primal-dual proximal point theory and Eular-Lagrange equation.And the projection of the dual variable can be deduced.Experiments show the proposed model can achieve satisfactory performance in terms of both subjective visual effects and objective evaluations.