Data Classification Based On Fisher Vector Coding And Sparse Constraint

With the development of technology in recent years,people use the Internet in their daily lives.This makes the image and data on the network grow dramatically.How to classify these images and data becomes a key issue in the field of intelligent information processing.Domestic and foreign researchers are very concerned about this area.Fisher vector coding has achieved remarkable results in image and data classification.In this paper,the Fisher vector coding algorithm is improved,and the sparse representation of Fisher vector coding is realized by means of optimization learning.Furthermore,the zero-shot learning classification algorithm is improved.Better classification results are achieved using these algorithms.The main results of this paper are as follows:1.A Fisher vector classification method based on optimized learning is proposed.The method uses a single Gaussian model to train the training samples at a time.By comparing the classification results and the real labels,the ‘better' single Gaussian models are selected.Based on these selected Gaussian models,the mixed Gaussian model is constructed to obtain the Fisher vector.Finally,Fisher vector is obtained.The method reduces the number of Gaussian models in the mixed Gaussian model and the coding complexity,which enhances the typicality of the feature and improves the classification accuracy of the data.2.A Fisher vector classification method based on semi-negative constrained sparse coding is proposed.This method constructs a sparse coding model,which effectively reduces the dimensionality of high dimensional Fisher vector coding,and removes the redundant components.In the sparse coding model,the weight of the coding is non-negative,so the semi-negative decomposition method is used to solve the sparse coding model.The feature code extracted by the method is more favorable for classification,and the efficiency of classification is improved.3.A classification method of zero-shot learning based on constrained embedded model is proposed.The method is used to optimize the embedding zero-shot learning classification method,and the constraints of L2 norm and L1 norm are adopted respectively.The L2 norm constraint prevents the over-fitting of the parameters.The parameter term of L2 norm constraint is solved by stochastic gradient descent method.The L1 norm constraint achieves the sparse representation of parameters.The parameters of L1 norm constraint are optimized by means of coordinate descending method.The two kinds of constraint strategies can realize the rationalization of classification parameters,so as to effectively improve the classification efficiency.