Research On Subspace-Based Cauchy Estimator Discriminant Algorithm

With the rapid development of artificial intelligence,image recognition problem in computer vision has become a hot topic,and subspace-based learning method plays a decisive role in the field of image recognition.Subspace learning algorithm is a classical dimensionality reduction method,which is based on the projection function to map the feature from the high dimensional space to the low dimensional subspace.Although the existing subspace-based learning methods have achieved comparable performance in the field of image recognition,there is progress in some cases of space.For instance,when there are confusing samples in the dataset and dataset contaminated by noise,the traditional subspace learning method will decline the recognition rate significantly.Therefore,for more effective dimensionality reduction algorithm needs further exploration and research.In order to solve the above problems,this thesis puts forward two kinds of dimensionality reduction algorithms of manifold learning based on subspace specifically,Cauchy Estimator Discriminant Analysis(CEDA)and Cauchy Estimator Discriminant Learning(CEDL)were used to solve the datasets contains confusing samples and the dataset contaminated by noise.The main work of this thesis:(1)This thesis introduces the background of subspace learning,expounds the development process and the present situation of the subspace,and analyzes the advantages and disadvantages of the commonly used evaluation function,and puts forward a more effective evaluation function to deal with specific problems.(2)The dataset mixed with confusable samples,using Patch Alignment Framework(PAF)to find the essential dimension of dataset,while using Cauchy Estimate theory to accurate modeling of confusing sample.(3)For the case where the dataset are mixed with noise,applying the Cauchy estimation theory directly alignment from local to global under the PAF,and use the regularization method to avoid overfitting.(4)The algorithms are integrated into the manifold learning framework,to facilitate the understanding of the algorithm and upgrade.(5)CEDA and CEDL algorithm are applied to face recognition and scene recognition experiments to illustrate the efficiency and robustness of the algorithm.Finally,the thesis analyzes and summarizes the research goal and the new algorithm,combined with deficiencies of the algorithm and dataset,then planning for the future improvement.