Pattern Classification Technique Based on Spatial Graph Distribution of Multiple Observation Samples
|Keywords||multiple observation samples label propagation joint sparse representation graphical presentation sparseness constructed graph distributed consensus reliability weighted inconsistent similarity measure|
Traditional pattern recognition mainly classify the single observation sample, whichis to say, the test pattern in each classification task only considers the image data of asingle sample. However, with the rapid advance of artificial intelligence technology, thetask of data acquisition is becoming easy，then the relevant data of multiple imagesbelonged to a specific pattern are captured easily, which are the multiple observationsamples. Multiple observation samples can provide more discriminatory information aboutthe specific pattern than single observation sample, by contrast, the classificationtechniques based on them have higher recognition rate. In view of this, on the basis ofanalyzing the related domestic and international research results, this paper studies themultiple observation samples classification algorithm.Firstly, exploiting the manifold structure of multiple observation samples, a multipleobservation samples classification algorithm based on graphical presentation ofinconsistent similarity measure graph is presented. First of all, considering the charactersof global and local data structure comprehensively, an inconsistent similarity measure isconstructed. The second step is to obtain the similarity matrix via inconsistent similaritymeasure graph, after that, the computation of the optimal map is transformed into thesearch problem of the largest eigenvectors of the Rayleigh quotient by a combinedGrassmannian kernel and then the projection matrix is obtained. Lastly, points on themanifold can be mapped into another space, the final classification is completed exploitsthe nearest neighbor classifier.Secondly, exploiting the distribution of the sensors network, a distributedclassification algorithm of multiple observation samples by sparseness constructed graphis presented. First of all, based on sparse representation to construct similarity graph andthen obtains the similarity matrix between samples. After that, utilizing the sensor graphand data graph, the label matrix and the objective function which captures the smoothnessof candidate labels are constructed formally on the basis of label propagation anddistributed consensus principle. Lastly, the estimation of the label of the test samples is completed by the optimization procedure of the objective function.Finally, exploiting the respective reliability of each single sample come from themultiple observation samples, a multiple observation samples classification algorithmbased on single sample weighted joint sparse representation is proposed in this paper.Multiple observation samples are first divided into single samples and each single sampleis processed separately by a sparse solution, obtained its respective sparsity and residual,then use them jointly to determine its reliability. After that, each single sample is weightedby corresponding reliability, the weighted multiple observation samples are reconstructed.Lastly, the classification is completed by sparse representation a second time.