Research Of Classification Method On High-Dimensional Image Data

With the developmet of information technology,the information collecting methods and techniques are more and more powerful.The high-dimensional data which are captured by human is increased rapidly in an exponential form.More and more uncertain data and mass fuzzy data have shown their high-dimensional properities.And high-dimensional data occurs at every corners of people's life and work.For example,with the increase of camera pixel-s,the resolution of images are higher and higher which causes the increase of dimensions of images.High-resolution images have improve the performances of most applications,such as image recognition,and added more burden for the whole image processing system.Therefore,how to deal with high-dimensional data and classify these data exactly has been a hot research topic in machine learning and data mining.This thesis mainly focus on classification for high-dimensional data.For the whole process of classification on high-dimensional data,this thesis utilizes dimension reduction technique to excavate important information and reduce the dimen-sions of high-dimensional data.Then in this thesis,I develop a new distance metric algorithm to measure the similarities between high-dimensional samples appropriately.Furthermore,for the diversity of various high-dimensional data,this thesis aims to introduce multi-view learning the-ory to extend some dimension reduction techniques to the multi-view framework.Finally,based on the methods proposed by this thesis,I utilizes deep learning to finish the task of fine-grained car classification,which can facilitate the various aspects of all people's life and work.In order to deal with a series of hard problems for classification on high-dimensional data,this thesis focuses on dimension reduction,distance metric,multi-view learning,deep learning respectively and achieve the following results.(1)Faced with the classification on high-dimensional data,dimension reduction should be considered firstly to improve the system speed and classification accuracy.Because some traditional sparse subspace learning methods maintained some incorrect similarities between samples,this thesis introduces a new sparse subspace learning method named locality structured sparsity preserving embedding(LSPE).Firstly,LSPE combines locality structured information with sparse representation to construct sparse correlations between high-dimensional samples.This procedure avoids the incorrect similarities casued by the global property of sparse represen-tation.Then,LSPE constructs the low-dimensional representations by maintaining the sparse correlations which has been calculated.Finally,because LSPE is an unsupervised subspace learning method.In order to deal with unlabeled samples and labeled samples simultaneous-ly,I extend LSPE to the semi-supervised framework in this thesis,which can obtained a better sparse subspace.The proposed dimension reduction method can obtain the low-dimensional representations for high-dimensional data and achieve better classfication accuracies.It can be utilized for the preprocess of high-dimensional data and facilitate the further researches on the later distance metric learning methods.(2)After the low-dimensional representations are obtained,in order to achieve better per-formances for high-dimensional data classification,learning an appropriate distance metric is of vital improtance for the whole system.Faced with the problem of semantic gap,this thesis introduced a method named semantic discriminative metric leaming(SDML)which can measure similarities between images exactly.Firstly,SDML combines geometric mean with normalized divergence,which separates images from different classes balanced in the metric space.And this procedure avoids the problem of semantic gap efficiently.Then,SDML adopts the large-margin criterion to separate images from different classes as much as possible,which can increase the discriminative ability of SDML greatly.Finally,SDML adopts 2 constrains to ensure an effec-tive and feasible solution.SDML aims to solve the problem of semantic gap and constructs an efficient distance metric,which has obtain good performances in the field of image classifica-tion.The proposed method can measure the similarities between samples exactly and improve the classfication accuracies of high-dimesional data.(3)Even though dimension reduction and distance metric can deal with high-dimension data classification from different viewpoint to improve the processing speed and classification accuracy,it's a hard task to deal with features from multiple views.Nowdays,one sample can always be represented by various high-dimensional features.In order to deal with multi-view high-dimensional data,this thesis introduced two new methods named co-regularized multi-view sparse reconstruction embedding(CMSRE)and multi-view sparsity preserving pro-jection(MvSPP),which can fully comprehend abundant information from multi-view features.CMSRE fully exploits the sparse reconstruction correlations between samples from multiple views and maintains these correlations into subspace.Through co-regularized scheme,CM-SRE leads features from mutiple views to learn from each other,which maintains the sparse recostruction correlations between samples from different views and construct a common sub-space for various views.Finally,CMSRE constructs an iterative procedure to obtain the optimal low-dimensional representations for multiple views.MvSPP projects multi-view features into kernel spaces to avoid the problem that multi-view features are from different spaces with differ-ent dimensions.MvSPP constructs different projection matrices for multiple views and projects multi-view features into one common space with a same dimension.This thesis further extends dimension reduction methods to multi-view framework to improve the classification accuracies.(4)With the development of large-scale image classification,how to make full use of re-lated techniques to deal with realistic problems is of vital importance but chanllenging.And recognizing the models of cars in highway is crucial for criminal investigation and highway control.This thesis aims to introduce a new method to recognize the models of various cars in highway and proposes a multi-path DCNN(MP-DCNN)model for this goal.Firstly,I utilize object detection models in this thesis,such as single shot multibox detector,to detect the whole cars in the images captured in highway.This procedure removes the useless background of cars.Then,multiple important parts of cars are detected,including car front,car logo and the whole car.Then,MP-DCNN is utilized to learn multiple parts from cars simultaneously.Finally,var-ious images are tested in the trained MP-DCNN model to verify the excellent performances of this car recognition system.This thesis focuses on a series of hard problems of the classification on high-dimensional data.After dimension reduction,this thesis introduces a new distance metric leanring method to measure similarities between samples,which can improve the classification accuracies.Further-more,due to the various representations of high-dimensional data,this thesis extends some di-mension reduction techniques to multi-view framework.Finally,this thesis utilizes the proposed methods and deep learning to finish the goal of fine-grained car classification.This procedure facilitate various aspects of people's life and work.