Research On Maximum Margin Classification Theory And Its Application

Maximum margin classification theory is an important topic in the field ofmachine learning in recent twenty years. In recent years support vector machine basedon the maximum margin classification algorithm is widely used in different fields, andhas been given more and more attention. Compared with the existing machine learningmethods, the maximum margin theory has a solid theoretical foundation and showexcellent performance in practice. This paper is based on the maximum marginclassification theory, its application in many aspects are studied, the main work is asfollows:The first, in order to improve the performance of image fusion, the traditionalalgorithm based on support vector transform (SVT) can not extract the directioninformation and Gauss kernel function not complete defects, puts forward a newmethod of image fusion of SVT and complex wavelet transform combining. In order toeffectively extract image information, the image using SVT transform orthogonalwavelet kernel decomposition, and then adopts dual complex wavelet transform(DTCWT) and high frequency information of the SVT transform decomposition, usingPulse Coupled Neural Network (PCNN) fusion strategy for image fusion. The relatedexperiment results and visual results show that the proposed method outperforms theexisting methods.The second, in the classification learning field, the support vector machine andtransfer learning combining with sample, in rare cases, guidance on the classificationmodel training the model using other existing related environment. Among them,especially the posterior probability is expressed in the form of classification results ofsupport vector machine classification results, to make up for the traditional supportvector machine0-1cannot well reflect the degree of defect important sample to themodel training and using a consistent regularization transfer learning framework for thedesign of the transfer learning algorithm based on support vector machine. The relatedexperimental results to achieve good, can further enhance the protection of data security,privacy by algorithm of distributed.The third, according to the classical minimax probability machine classificationusing off-line learning algorithm (batch), involves much data, using two order coneprogramming is convex optimization problems, can not be caused by the large amount of calculation meet the increment or decrement of learning, timely classificationproblems etc., put forward the increment decrement the minimax probability machinethe learning algorithm. The algorithm can be performed on the data processingsequence, but is not limited to batch, meet the increment decrement learningrequirements. Experiments show that, the algorithm has low computational complexity,approximate real-time etc..