Research On Co-training Algorithm And Its Application In Classification

With the rapid development of computer technology and wide application of digital electronic devices,it is easy to obtain a massive amount of multimedia data including pictures or video.It is a great challenge to process these huge multimedia data described by multiple features,especially where most of the data are unlabeled or annotated in real world application.Co-training is one of the important semi-supervised learning algorithms that can tackle the above mentioned problems.It works in an iterative manner to combine two view representations,in particular,two classifiers are first trained on two different views,then teach each other and update to improve recognition accuracy.Co-training has attracted much attentions and many variants of co-training have been reported in natural language processing and pattern recognition and other fields.The main contributions of this thesis are as follows:1.Hessian regularized co-training is proposed.It is easy to introduce noise since the initial classifier has only mediocre accuracy,to address the problem,Hessian regularized co-training is proposed which combine Hessian regularization algorithm to co-training.Hessian can employ a large number of unlabeled samples information to exploit the local structure of underlying data manifold,and obtain satisfying results in scene recognition.2.A general framework of co-training is proposed.Since many co-training variants and applications have been developed,it is essential and informative to provide a systematic framework for better understanding these algorithms.We summarize the related co-training variants,and divide into three categories: co-training on multiple views,co-training on multiple classifiers,and co-training on multiple manifolds,and obtain satisfying results in action recognition.3.Co-trained density spectral clustering is proposed.We successfully introduce the idea of co-training to spectral clustering,which is belong to unsupervised learning.In this paper,we construct multi-view spectral clustering model,and apply it to action recognition,image Identification,greatly improving the accuracy of clustering.