Image Inpainting Based On Belief Propagation Model

Since the Priority-BP algorithm can restore not only damaged texture region but also structural region,it has the incomparable advantage comparing with other algorithms.Furthermore,it also has a good performance to inpaint such region with damaged texture and structural information simultaneously.However,Priority-BP algorithm is low efficient so that it cannot satisfy real-time requirement.To solve this problem,three improvements are proposed in the paper.1)Before the first iteration,the Priority-BP algorithm has high computationally expensive in nodes interaction procedure because the internal nodes lack prior information.In our approach,the dictionary representation is used to prune labels before message propagation among nodes in order to cut computation and accelerate convergence.Firstly,the damaged images are divided into two types:known blocks in source region,and restorated blocks in damaged region.Secondly,the fixed dictionary is constructed based on known blocks,and its dictionary coefficients are also initialized.Then,these optimal approximation coefficients are obtained by using known information in restorated blocks.Finally,the rough information of restorated blocks can be obtained based on this dictionary and its optimal coefficients and be used as reference information for the internal node.2)To prune redundancy labels of Priority-BP algorithm,label clustering based on SSD is constructed in our approach to improve interaction efficiency among nodes.Firstly,the active labels are ranked according to the confidence value,and the label with greatest confidence is selected as the first cluster center.Then,similarities between this first cluster center and its subsequent labels.If they are sufficiently similar,the subsequent label will be pruned.Otherwise,the subsequent label will become new cluster center.This iteration can gradually prune all redundant similar labels of one node to improve the compactness label sets.3)In our approach,the priority computation is improved to facilitate label clustering.A new structural term is introduced into the calculation formula for priority,which can find special node by using known images information of its neighborhood to improve its priority.This term can compensate priority perturbation caused by label clustering in order to guarantee structure information repaired correctly.In summary,our improvement algorithm can reduce computation cost of message interactions among internal nodes and prune redundancy labels of every node.The experimental results show that our improvement algorithm can accelerate information propagation,reduce convergence time,and improves execution efficiency.