Pattern Classification Technique Based On Sparse Representation Optimization Constraints Of Multiple Observation Samples

Traditional pattern recognition is often only for classifying the single observationsample of a certain domain, namely training samples and test samples in eachclassification task are from the same domain, and test pattern has just single observationsample. However, with the development of information technology, which leads to thevisual data of the same or different domains increasing dramatically, then multipleobservation data belonged to a specific pattern are captured easily, which are the multipleobservation samples. Multiple observation samples of the same or different domains canprovide more discriminatory information about the specific test pattern than singleobservation sample of a certain domain, then the classification techniques based on themhave higher recognition rate. As a result, on the basis of the domestic and internationallatest related research results, this paper studies the multiple observation samplesclassification algorithm.Firstly, exploiting similarities and differences of multiple observation samples, amultiple observation sets classification algorithm based on joint dynamic sparerepresentation of low-rank decomposition is presented. At the beginning of finding thebest set of image transform domain, which decompose the data matrix into a low-rankmatrix and an associated sparse error matrix. After that, the low-rank matrix and sparseerror matrix is represented by joint dynamic sparse respectively. Lastly, we compare theclassification results with the total sparse reconstruction errors.Secondly, exploiting transfer matrix model, a multiple observation sets classificationalgorithm based on sparse representation of transfer learning is presented. First of all,adapts object models acquired in a particular visual domain to new imaging conditions bylearning a transformation that minimizes the effect of domain-induced changes in thefeature distribution. After that, the transformation is learned in a supervised manner andcan be applied to categories for which there are no labeled examples in the related domain;Lastly, we compare the classification results with the sparse reconstruction errors.Finally, exploiting the framework of nonnegative matrix tri-factorization, a multiple observation sets classification algorithm based on transfer learning of L1-Graph ispresented. First of all, basing on transfer learning to build a framework of nonnegativematrix tri-factorization, we exploit these unchanged information as the bridge ofknowledge transformation from the source domain to the target domain; The second stepis to construct L1-Graph on the basis of sparse representation, so as to preserve thegeometric structure of samples and features; Lastly, exploiting an iterative algorithm tosolve optimization problem, and then the estimation of the label of the test samples iscompleted.