Data Classification Based On Sparse Fisher Vector Coding

In recent years,with the fusion development of the computer and communication industry,internet terminals,such as mobile phone and tablet computer and so on,are extensively applied,which generate explosive data growth.Therefore,as an important subject of data mining,data classification has been concerned by researchers at home and abroad.The extraction of the global feature based on local feature is the common step for data classification.However,compared with other methods,the classification performance based on Fisher vector encoding(a global feature)is better than that of the traditional Bag of Words(Bo W)model.In this paper,we focus on the data classification based on Fisher vector encoding.The main research results are as follows:1.A sparse representation classification algorithm based on Fisher vector encoding is proposed.The method removes the redundant components in Fisher vector information using sparse coding.At the same time,in the process of Fisher vector encoding algorithm,for extracting characteristic coding representation,a locally constraint factor is introduced,which achieves the efficient coding and reduces the coding error.In the method,the efficient information components of data is extracted,which is beneficial to improving classification performance.2.A Fisher vector sparse encoding classification algorithm based on Boosting is proposed.The method utilizes Boosting to rank the Gaussian kernel and to selecte the number of Gaussian kernel.By selecting the Gaussian kernel,the Fisher vector encoding is optimized.At the same time,this paper achieves an important step to adjust the weights of the Gaussian mixture model after selecting Gaussian model using Boosting method.By normalizing the weights of the Gaussian mixture model,the proposed mothed further enhances the information contained in the sparse Fisher vector and improves the accuracy of the data classification.In this thesis,we present two kinds of sparse Fisher vector encoding algorithms,which are tested in the five databases.The experimental results show that the two Fisher sparse algorithms have better classification performance than the related algorithm.