Research On Multivariate Compression Technology For Volume Data

Recently, with the rapidly development of mobile computing and sensor equipment,multidimensional and multivariate data information is increasing. For example, a company calledAlibaba, which is one of the giant of internet enterprise, that it has focused on the electroniccommerce for many years, and collected a huge mass of text and video data in its database. Amongthem, for the great part of is multidimensional multivariate business data which have hugecommercial value. In order to make a more reasonable analysis of these massive datasets,compression technology certainly be the one of the key role. Data compression can remove a greatdeal of redundant information to reduce the pressure of data mining and visual analysis or otherapplication. Traditional single variable data compression techniques can’t do well with suchmassive data. However, there are little compression algorithms which can process with large-scalemultidimensional multivariate datasets. Therefore, multidimensional and multivariate datacompression technology contains enormous research value.Aim at this kind of data. This thesis analyzes the relation of multiple attributes of themultivariate volume data and proposed two kinds of different compression methods which canbring efficient and feasible results.In the third chapter, we propose a multivariate volume data compression algorithm based onthree-dimensional wavelet transform. The most important advantage of using wavelet transform isthe local properties in the time domain and frequency domain. In addition, we can combine it withmulti-resolution analysis to get more retails. Firstly, we transform multivariate volume data fromRGB color space to YCbCr color space, and then according to the different weight value of threecolor components uneven sampling. Secondly, we use wavelet transform and discrete cosinetransform to decomposition different color components. At last, we set a threshold to removeredundant information to implement compress multivariate volume data.Considering the compression algorithm in the third chapter only can process with color volumedata. In the fourth chapter, we propose another innovative compression algorithm which combinedwith some mainstream theory in machine learning field. Considering the manifold structure ofmultivariate volume data field, we could select the most representative feature vector sets byvariance minimization. The compression results are consist of the representative feature vector andthe volume data which have being reduced the dimension by MDS. After all, we could reconstructthe volume data’s multiple variables by semi-supervised learning algorithm. Based on therepresentative feature vector sets, the way is that we could learn a regression model then can use it to forecast the other variables.