Least Square Based Feature Extraction

During the Internet era,high-dimensional datasets emerging in large numbers have become a great challenge in pattern recognition.Therefore dimension reduction techniques have served as fundamental ways to processing high-dimensional data and dealing with ' the curse of dimension'problem.Researches on dimension reduction have got a certain achievements in recent years.Due to the size and dimension of high-dimensional data,extracting data features is tough work;nevertheless it has good prospects for the future.This paper studies several classical dimension reduction algorithms.Inspired by these methods,we work out some new algorithms to optimize the efficiency and processing speed.The followings are the main work and contributions of this paper:1.This paper makes a brief introduction and summary of the methods currently used in dimension reduction,and uses a general graph embedding framework to uniformly represent these methods above.Aiming at the disadvantage of slow computing speed in the traditional graph embedding framework,we propose a Regularized Least Square based Graph Embedding framework(RLSGE),which combines the advantage of sparse coding theory and least square procedure.RLSGE optimizes the graph construction mode.It speeds up dimension reduction procedure while maintaining the robustness of original graph embedding framework.Under the RLSGE framework,an improved L2-graph based dimension reduction method is proposed:Regularized Least Square based Discriminative Projections(RLSDP).As long as inheriting the sparsity characteristics of SPP,RLSDP optimizes the speed of graph constructing.Moreover,it turns the unsupervised leaning mode into supervised leaning,which therefore well adopts the label information on sample data,and improves classification accuracy.2.In the area of action recognition,we survey feature extraction methods.Combined with time-domain information in video clips,we propose a partial-least-square-based dimension reduction descriptor for classification.We extract the original HOG feature,construct space-time gradient orientations merged with spatiotemporal information,and execute fast dimension reduction with partial least square method.This method utilizes class label information in data features.Experiments on benchmark datasets show the improved efficiency of proposed method.