Font Size: a A A

Multiple Observation Samples Classifier Algorithm Of Space Distribution Association Description

Posted on:2013-01-12Degree:MasterType:Thesis
Country:ChinaCandidate:L L WangFull Text:PDF
GTID:2218330362963084Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
The classification problem solved by the traditional pattern recognition is often onlyfor the case that test pattern has just single observation sample. However, with the rapiddevelopment of information technology, the task of data acquisition has become more andmore easy. The amount of data that can be obtained by people is more and more large, andparticular, it commonly happens that multiple observation samples of a pattern arecaptured. Multiple observation samples can provide more information about test patternthan single observation sample, so as to increase classification accuracy.Firstly, the samples in each class set can be supposed to distribute on a same lowdimensional manifold of the high dimensional observation space. With regard to how totake advantage of this manifold structure for the effective classification of the multipleobservation samples, label propagation classification algorithm of multiple observationsamples based on L1-Graph representation is proposed. Based on sparse representation toconstruct L1-Graph. Then, on the basis of semi-supervised label propagation algorithm totransform the computation of the optimization label matrix to an optimization problem ofdiscrete object function and obtains the class of the test samples.Then, use the convex hull model to approximate the sample distribution of highdimensional and transform the problem of classification of multiple observation samplesto the similarity of convex hulls. The classification algorithm of multiple observationsamples based on L1norm convex hull data description is proposed. Construct the convexhull for each class in the train set and multiple observation samples in the test set as thefirst step. Then determine the distance between convex hulls is zero or not, use L1normdistance measure to solve the similarity of convex hulls or the reduced convex hulls. Thenearest neighbor classifier is used for the classification of multiple observation samples.Lastly, use the similarity of subspace, the classification algorithm of multipleobservation samples based on kernel discriminant canonical correlation (KDCC) isproposed. The original input space is nonlinear mapping into a high dimensional featurespace, use an iterative algorithm to train the optimal KDCC matrix, and then obtain the transformed kernel subspaces. Then the nearest neighbor classifier is used to solve theclassification of multiple observation samples.
Keywords/Search Tags:multiple observation samples, L1-Graph, label propagation, convex hull, L1norm distance measure, canonical correlation, kernel discriminant canonicalcorrelation, nearest neighbor classification
PDF Full Text Request
Related items