Research Of Depth MAP Inpainting Algorithm Based On Depth Edge Discrimination Model

As an important way to represent 3D information,depth maps have been widely used in computer vision systems.However,depth maps are not perfect because of sensors and environmental factors,they may suffer from defects such as low resolution,unknown regions,and noise.Among these problems,unknown regions is a difficult issue.Therefore,this thesis researches on unknown regions inpainting in a detective depth map base,hoping to put forward a more effective algorithm.At the same time,this unknown regions inpainting problem which is usually called hole-fill problem in depth maps.The main work of this thesis can be summarized as the following aspects:(1)Through the study of common depth image inpainting algorithms,it is found that the object edges extracted from depth maps(which called depth-edge)can be effectively used in local-based algorithm,these edges can be guided information in inpainting algorithm.But unfortunately,as the unknown regions in depth map,algorithm only can extract object edges(which called color-edge)from color image.Therefore,this thesis research on how to extract useful depth-edge and how depth-edges works.(2)we proposed a framework for unknow regions inpainting in depth map,which based on depth-edge discriminant model.The framework consists of two parts: the first part is an edge extraction module,which actually is a depth-edge discriminant model,including a depth-edge probability extraction model and a candidate-point extraction algorithm.Firstly,a set of candidate-points are extracted by edge operator in color image.Then,module selects depth-edge point from the candidate-points set based on depth-edge probability extraction model(e.g.convolutional neural network model);The second part is a hole-fill module,which consist of Joint-Trilateral Filter,Directional Joint Bilateral Filter,and Partial Directional Joint Bilateral Filter.It fill the unknown regions in depth map according to depth-edge information.(3)In order to validate the performance of this framework,we test on the Middlebury Stereo dataset,the result shows that: firstly,Directional Joint Bilateral Filter is better than Partial Directional Joint Bilateral Filter in edge region;secondly,Joint-Trilateral Filter perform well in noise processing;finally,the framework has a certain improvement in inpainting effect on the data,compared with common algorithms.