Font Size: a A A

Research And Application Of Image Classification Algorithm Based On Semi-supervised Learning

Posted on:2018-08-16Degree:MasterType:Thesis
Country:ChinaCandidate:X TianFull Text:PDF
GTID:2358330542462929Subject:Software Engineering Theory
Abstract/Summary:PDF Full Text Request
Machine learning is to make the machine can learn like human beings,through the study of some observation samples,find the hidden law of the sample,and then establish some models to identify the unknown samples.The traditional machine learning includes supervised learning and unsupervised learning,but both have some limitations,for supervised learning,acquiring more time and effort to get labeled samples.For unsupervised learning,without labeled samples to guide the training,the result of training is not unsatisfying.Semi-supervised learning combines the advantages of supervised learning and unsupervised learning,fit for image classification problems.Under the guidance of a small number of labeled samples,using a large numbers of unlabeled samples to train the classifier,improve the accuracy of classification and generalization ability of classifier.Label mean semi-supervised support vector machine is one of semi-supervised classification branches based on discriminative,take the mean of unlabeled samples as a constraint,simplified the traditional semi-supervised support vector machine to solve complex problems.Label propagation through minimax paths algorithm belongs to the category of graph based semi-supervised classification,build a sparse neighborhood graph,spread class labels by minimax path,reduce the label propagation algorithm's time.But these two algorithms still exist the following problems:for the mean labels,image spectral features is not considered completely,so accuracy of classification is not high.For the label propagation through minimax paths algorithm,it is necessary to construct a sparse nearest neighbor matrix.Inappropriate K value can result in disconnecting and make some data without a label.In this paper,according to these two issues,the specific work is as follows:(1)Semi-supervised support vector machine using label mean(meanS3VM)will lead to low classification accurate and unstable result due to random selecting unlabeled samples.In order to deal with these problems,a semi-supervised support vector machine based on clustering label mean image classification algorithm is proposed.This method modifies the penalty term of the original algorithm for unlabeled samples,clusters unlabeled samples and replaces label mean by clustering label mean.The experimental results indicate that the proposed method greatly reduces the misclassification of background and objectives,improves the stability and accuracy of the classification algorithm.It is suitable for image classification.(2)Label Propagation through Minimax Paths is a semi-supervised classification method,which has the advantage of low time complexity.The method requires the sparse similarity matrix constructed by K-NN graph.For different data,how to determine the value is a problem.Inappropriate K value can result in disconnecting and make some data without a label.This paper presents a method of Label Propagation through Adaptive Nearest Neighbor based on Minimax Paths.For image classification problem,this method can calculate the adaptive nearest neighbors of each sample point,and solve the problem of disconnecting which caused by inappropriate K value meanwhile improve the classification accuracy of the algorithm.(3)Apply the Label Propagation through Minimax Paths algorithm and the Label Propagation through Adaptive Nearest Neighbor based on Minimax Paths to hyperspectral remote sensing image classification.Combined with the low rank recovery technique,improve the classification accuracy of the two algorithms.Apply the Label Propagation through Minimax Paths algorithm and the Label Propagation through Adaptive Nearest Neighbor based on Minimax Paths are applied to multispectral remote sensing image classification.Combined with the meanShift smoothing algorithm,improve the classification accuracy of the two algorithms.
Keywords/Search Tags:semi-supervised learning, semi-supervised support vector machine(S3VM), label mean, sparse matrix, label propagation
PDF Full Text Request
Related items