Research On The Key Technologies Of Visualization Analysis For Volume Data

Volume data generally refer to discrete data sets distributed in three-dimensional space and interacting with each other.They are multimode,multidimensional,multivariate,and large-scale.Volume data,such as CT,ocean flow field,nuclear explosion,and hurricane data,are widely produced in scientific and engineering domains of medical,fluid dynamics,nuclear physics simulation,and meteorological science,etc.,and the processing and analysis of volume data are of great significance for medical diagnosis,maritime navigation,weather forecasting,and military applications.Visualization analysis is a kind of analysis method with the help of visualization,by presenting the abstract volume data with concrete graphs and images,people could intuitively explore and analyze the hidden patterns,relationships,and features within the volume data based on visual perception,and further get an in-depth understanding of the internal mechanism and changing rules of the related physical phenomena.Currently,the visualization and analysis of volume data are suffering from severe challenges: first,multimodality leads to information loss in the data acquisition and application,resulting in the lack of completeness and validity;second,large scale leads to high hardware requirements,massive visual mapping,and heavy overlap,resulting in low efficiency and quality of the visualization and analysis;third,multiple variables leads to complex interactions within the volume data so that it is difficult to explore the relationships among variables.Since data processing and visual analysis are two important parts in the visualization analysis of volume data,this dissertation researches the key problems including interpolation,segmentation,feature extraction and tracking for time-varying data,and correlation analysis for multivariate data,expecting to break through the key technologies and lay a theoretical and technical foundation for the visualization analysis of volume data.The main works of this dissertation are as follows:1.This paper proposes an interpolation approach based on statistical distribution and sampling technology that integrates the spatial location dependency with the local statistical property,which aims at solving the problems of lack of scientific guidance in selecting the reference samples and insufficient utilization of the data information of the inverse distance weighted method.First,we search the sample set based on the distance impact factor.Second,we model the sample set whose probabilistic distribution function(PDF)is estimated by the improved Gaussian mixture model(IGMM)which truncates the tail probability.Then we optimize the IGMM based on Bayesian information criteria.Finally,we estimate the interpolation value by performing sampling on the IGMM.Experimental results show that this method is of high accuracy and strong robustness.2.This paper designs a strategy of “over-segmentation and clustering” and proposes a hybrid segmentation method based on supervoxel clustering,which aims at solving the problems of poor generality,and weak balance between accuracy and efficiency.First,we propose the distance weight and attribute weight to optimize the criteria in region growing,and over-segment the volume data to generate supervoxels based on the standard region growing algorithm.Then we introduce the information diffusion estimation(IDE)and optimize the diffusion estimation coefficient,and we define the statistical distance based on the IDE.Finally,the supervoxels are clustered by the K-means algorithm,and the volume data are segmented.Experimental results show that this method has higher time efficiency while ensuring the segmentation accuracy.3.This paper proposes a similarity matching of statistical property guided feature extraction and tracking method that employs the statistical distribution as the feature description,which aims at solving the poor generality of the existing methods that heavily rely upon feature definitions.For the feature extraction,we firstly generate supervoxels based on the simple linear iterative clustering method.Then we use a hybrid modeling strategy to establish the statistical description and representation of the volume data based on the PDF.Finally,we interactively select the reference feature,and then,the statistical distance between it and each supervoxel is measured,further,the feature is extracted by matching the similarities.For the feature tracking,we regularly partitioning the volume data to generate supervoxels and update the statistical distribution parameters based on an incremental learning scheme,then the feature tracking is implemented by feature extraction step by step.Experimental results show that this method has high accuracy and efficiency,which is independent of the feature definition,and is suitable for multiple features.4.This paper proposes a visualization-based approach for correlation analysis,which aims at exploring the internal relationships among the variables.First,we select the most representative variable as the reference variable,and discretize all the variables based on the histogram.Second,we analyze the relationships among the components of the reference variable and the other variables based on information overlap,and the associations among all the components based on information flow,respectively.And then,we construct a surprise-influence map to guide users in the identification of the representative components of different variables.Finally,we explore the correlation among variables by visualizing these identified components and analyzing their interactions.Experimental results verify the applicability and effectiveness of this method.