Research On 3D Data Segmentation Method Based On Supervoxel

With the improvement of computer science technology and applied mathematical theory,graphic image segmentation technology has become a key preparation step from low-level image processing to high-level object recognition and understanding,which has been widely used in medical image processing,computer 3D animation,man-machine interface,intelligent virtual reality,computational simulation visualization and many other application areas.3D model data is a new form after sound,image,video,and the segmentation of 3D data has always been a hot topic in the fields of pattern recognition and computer intelligence,which has been defined the division of logically unified 3D data into small,independently manageable physical units,effectively improving the efficiency of index and sequential scanning,and convenient for subsequent refactoring,reorganization and recovery.People extend the conception of image superpixel to 3D space supervoxel,simultaneously,the research work of combining supervoxels with the processing and analysis of 3D data have been become a research hotspot in the computer graphics.Supervoxel is a new image preprocessing technology that has emerged and developed rapidly in recent years,it characterizes voxel in local regions of volumetric data and aggregates similar or resemble voxel to generate a sub-regions,which has a stable local structure,similar local feature and actual local significance.Compared with the pixel or voxel,which is the basic unit of image graphics in the traditional segmentation technology,there are three main advantages of incorporating supervoxel.First,it greatly decreases the data size of the algorithm,and effectively reduces the complexity of subsequent processing and analysis work.Second,it is easier to extract local features,which lays a foundation for the subsequent image segmentation that can obtain structural information correctly.Third,the supervoxel characterizes a more smoother and softer description of the object boundary,which makes up for the inadequacy of pixel or voxel boundary description,and makes it easier to obtain accurate object boundary segmentation.The paper using the visual characteristics,carring out the feature extraction and similarity calculation of voxels and supervoxels,from three different forms of 3D data,namely,the methods of volumetric data segmentation,3D point cloud segmentation and RGB-D data segmentation are deeply studied and researched,exploring and researching a closed 3D volumetric data,or a 3D point cloud scene,or a group of RGB-D datas is divided into a group 3D data sub-blocks,which is not only connected to each other,but also has a certain number,and each has its own certain meaning,in the meanwhile,experimental verifications are given.The main work and innovations of this paper are as follows:Firstly,for the problem of medical volumetric data segmentation,an improved kernel-based 3D fuzzy C-means clustering method is proposed,which solving the segmentation of supervoxel clustering.In the generation of supervoxel,a method of selecting seed points is recommend,that is establishing the gray histogram,finding the point of the gradient minimum as the seed point in each histogram.Then the color distance and spatial distance between the voxels,and the coordinate difference between the voxel and the cluster center are calculated,using these cariables,Simple Linear Iterative Clustering(SLIC)method is extended to 3D and caused dividing the voxel image into supervoxels.Finally through improved kernel-based 3D fuzzy C-means clustering algorithm realizes clustering segmentation of supervoxels.By constructing the membership degree of spatial constraints,the kernel-based fuzzy C-means clustering is improved,which makes the results more stable when dealing with noise-based volume data segmentation,retains more details,and effectively improves the accuracy of 3D human brain tissue segmentation.The algorithm can process volumetric data and perform complete 3D volumetric data segmentation,the experimental samples prove that the algorithm is easy to handle a large number of special regions,which are often not solved by the classical fuzzy C-means clustering method,at the same time,the experimental results of the proposed algorithm applied to the human brain dataset are fully introduced and discussed in depth.Secondly,facing the point cloud data,the feature of supervoxel is represented precisely basing on the construction of local coordinate system,a new method of 3D point cloud scene segmentation is proposed,which increasing the description of the normal vector of center point in the local coordinate system,combining visual characteristics with supervoxels and using proximity,similarity and continuity to further describe supervoxel features,so that it can produce the segmentation results of meeting the human visual characteristics.The first step is to transform the whole point cloud into a 3D grid structure and to generate supervoxel by clustering voxel according to the space coordinate distance,the curvature value distance and the FPFH feature description.The second step is to build local coordinate system framework basing on point cloud density weight in the supervoxel,which makes the calculation of the normal vector of center point more accurate,adds the visual characteristics,proximity,similarity and continuity to calculate the geometric characteristics of the supervoxels.Finally,a graph model is constructed around each supervoxel,and graph-based clustering is applied to merge supervoxels according to their similarities,so that to produce the final segmentation result.In order to verify the effectiveness of the proposed algorithm,the generated supervoxel and final segmentation results are discussed respectively.The experimental results show that the proposed algorithm can realize meaningful and reasonable segmentation for complex 3D point cloud scenes,especially robust to the segmentation of object boundaries,and provide conditions for the increasingly mature model high-level semantic analysis.Thirdly,a new method of RGB-D data segmentation based on visual saliency map which is used to guide the generation of supervoxel is proposed.For the input RGB-D data,using Wavelet Transform visual saliency image detection method to generate the saliency maps,and basing the anisotropic center-surround difference figure to compute the depth saliency map,then fusion of the two ones.Next the corresponding point cloud data is generated for the input RGB-D.Later,the uninterested sampling points are selected as the seed points,and the average saliency distance,spatial feature,color feature and brightness feature distance metric are used to describe the point cloud voxel to generate the initial supervoxel.At last,basing on the supervoxel geometric distance and color information feature description,the final segmentation result is obtained by iteratively merging.Through experiment analysis,compared with the effectiveness of three typical visual saliency detection methods in the segmentation,they are single-used Wavelet Transform,Minimum Barrier Distance Transform and our chapter algorithm,it is proved that the segmentation result of visual saliency map is more accurate and uniform,and the calculation efficiency is more convenient and efficient.