Anti-Texture Cross And Weighted Cross-based Stereo Disparity Refinement

Stereo matching is a very popular discussion topic in computer vision,because this topic could get 3D information of the real world scene from a pair of color images and then a series of operations could be done.Stereo matching has been used in many applications,including autonomous vehicles,3-D scanning,3-D construction and 3-D tracking.All the algorithms in stereo matching are divided into global algorithms and local algorithms.Because of the low efficiency of global algorithms and the low computational complexity of local algorithms,the local algorithms are drawn more and more attention.In the study of local algorithms,many researchers have found that there are lots of occluded regions and highlighted regions in the color image which cause lots of unmatched pixels in the raw disparity maps.Disparity refinement is a significant step in stereo matching,which plays an important role.This step could identify and correct the error pixels in raw disparity map and generate a better disparity map.More and more researchers in stereo matching focus on this area for the reason that disparity refinement in stereo matching could improve the quality of disparity map a lot and simplify the computation.In the study of disparity refinement,almost all the local algorithms need to build proper support regions for the error pixels.These support regions should find similar pixels for the error pixels.The definition of similar pixels is that pixels with same disparity value in neighboring area or in the same plane in the real world scene.However,how to construct support region adaptively and identify the border of objects have been an important topic.In order to build support region adaptively,cross-based support region construction algorithm is widely used.Because the expansion of the cross depends on the color similarity,while the color intensity is fluctuate in the texture area.The traditional cross-based algorithm suffers a lot from the texture in the color image,which could impede the expansion of the cross.This paper proposed a disparity refinement algorithm based on two-stage cross.The first stage is anti-texture cross-based support region construction algorithm.In order to reduce the influence of texture,multi-layer structure extraction is proposed.Then anti-texture cross with exploration vector is used to build a proper support region for error pixels.According to these support regions,the second stage of the algorithm is proposed,which is called weighted cross-based updating algorithm.The Middleburry benchmark is used to do the final comparison.The experiments show that the proposed method could build support region accurately and improve the accuracy of the disparity map in the final results with fast speed,compared to other tree-based algorithms.It also outperforms the existing disparity refinement algorithms in preserving the boundaries of objects in the final disparity map.