Disparity Map Optimization Based On Sparse Subspace Clustering

Constrained and affected by hardware conditions and environmental factors,the disparity map generated in stereo matching may contain relatively serious noise,data defects,and occlusions.These disparity defect areas make it difficult for the disparity map to reflect the consecutive depth information of the scene.The disparity map optimization technology uses the disparity optimization algorithm to predict unreliable or missing disparity values and achieves the purpose of reconstructing an ideal disparity map under the condition that the disparity information is partially intact.As one of the important links of 3D reconstruction,this technology has attracted wide attention in the field of stereo vision and image processing in recent years.This thesis first introduces research background and significance,research progress at home and abroad,and problems in current research of disparity map optimization techniques and sparse subspace clustering.Next,this thesis describes the relationship and difference between disparity map optimization and image inpainting by combining the basic theory of ill-posed problems and ill-posed inverse problems.On the basis of these theories,this thesis focuses on the feasibility of natural image processing techniques in the optimization of disparity maps.Considering that the sparse and example-based natural image inpainting algorithm can largely complement the missing parallax information in a large area,it is less time-efficient and difficult to meet the application requirements of some real-time systems.This thesis proposes a disparity map optimization algorithm based on sparse subspace clustering.The confidence-item and edge structure are used to determine the priority of pixels to be completed.The intact pixel information in the disparity map is clustered and a multi-class dictionary is established.The sparse representation is used to reconstruct the disparity information,and the defect disparity map is optimized.In the simulation experiment of disparity map optimization,a multi-scene standard disparity map data set is cited in this thesis,and the proposed algorithm is compared with the traditional algorithms used in disparity map optimization and the representative algorithms in natural image inpainting algorithms.The mean square error,peak signal-to-noise ratio,and structural similarity index of the optimization results are counted and analyzed to compare the subjective and objective applicability of algorithms.The results show that the disparity map optimization algorithm based on sparse subspace clustering proposed in this thesis is not only more stable than the disparity map filtering method in processing disparity map point defects,but also can restore disparity information under a higher defect rate.When dealing with large-area irregular defects,the method in this thesis can effectively reduce excessive smoothness and has a higher degree of reduction of parallax compared with the traditional sparse and example-based image inpainting methods.