Research On Visual Pattern Recognition Based On Graphical Representataion Of Multivariate Data

Automatic pattern recognition (PR) is usually considered as an engineering area which focuses on the development and evaluation of systems that imitate or assist humans in their ability of recognizing patterns. It may, however, also be considered as a science that studies the faculty of human beings and possibly other biological systems to discover, distinguish, characterize patterns in their environment and accordingly identify new observations. The engineering approach to pattern recognition is in this view an attempt to build systems that simulate this phenomenon. By doing that, scientific understanding is gained of what is needed in order to recognize patterns. The paper mainly deals with weaker assumptions of the models, such as the issue of representation of data. Statistical pattern recognition traditionally relies on a feature representation. So based on the research of the graphical representation of multivariate data for many year, the paper proposed that the data feature representation in traditional pattern recognition was transformed the multivariate graphical representation of data. The novel research direction was opened, that is the pattern recognition based on graphical representation of multivariate data.The paper's background was the issue of representation and classification of data in pattern recognition. The paper's means was the 2D graphical representation of multivariate data. The paper's base was the graphical feature extractor and graphical classifier of the multivariate graph. The paper's aim was the establishment of the theoretical framework of the pattern recognition based on graphical representation of multivariate data. The visual pattern recognition novel method will be our ultimate goal.First, the general theory of the multivariate graphical representation of non-graphical feature data was proposed. The mathematic theoretical framework of representation (formula representation, matrix representation, geometric representation) was researched. The uniqueness theorem of multivariate graphical representation was proposed and proved.Secondly, the graphical feature extractor methods of the multivariate graph were researched. The geometric view of the multivariate graph was proposed. Especially, the area or barycenter graphical feature of the star plot graphical representation of multivariate data was proposed. The feature order issue of the multivariate graphical representation of non-graphical feature data shall result in that the classification performance should be different based on the generated graphical feature of the multivariate graph of different feature order. So the corresponding feature order menthods within the optimal classification performance were researched. The first method was the feature order of traditional feature selection. The second method was the feature order based on the stochastic search algorithm with global optimization (such as Genetic Algorithm et al). The Third method was the optimal feature order based on some certain criterion function fitted the issue.Thirdly, the idea was proposed and the method was practiced of the ascending dimensions transform and feature selection. The ascending dimensions transform and feature selection meaned that first the graphical feature of the multivariate graphs of all feature order were extracted, and then the part feature from the original feature and extracted graphical feature were selected. It guessed that the classification performance of the selected part feature was at least the classification performance of the distinguishing graphical feature derived from the optimal feature order. And then the experiential connection of the number of the selected part feature dimension and the classification performance was researched and used to guide the selected number.Last, the graphical classifier and graphical clustering were researched. The definition and formula of the graphical dissimilarity was proposed. The single prototype graphical classifier based the graphical dissimilarity and template matching was proposed. The K nearest neighbor graphical classifier based the graphical dissimilarity and K nearest neighbor was proposed. The graphical classifier and graphical clustering based on the Chernoff faces were also researched and was used in the mulitidimensional multiresource information fusion for Health Smart Home.The whole system was implemented using the Matlab programme. For some UCI real dataset such as wine, breast cancer and diabetes, the obtained best classification error of distinguishing graphical feature of the star plot is 0%,1.61% and 20.7%, which was very superior to the previously reported optimal classification performance of the international compared baseline methods. The experimental results proved our idea.