Research Of Image Classification Based On Bagging And Tri-Training Algorithm

With wide application of electronic digital equipment and the rapid development of Internet, especially the new rising social software(network) We-chat, QQ and micro-blog,the digital image information is growing quickly every day. Therefore, how to speed up,retrieve effectively, classify and dig out the useful information becomes a focus of current research, in which many researchers pay attention to the field of image classification.Machine learning is a procedure which can be transformed from raw data into useful information automatically. According to the different forms of learning, machine learning is divided into supervised learning and unsupervised learning. Semi-supervised learning has been put forward and attracted attentions since 1980 s, in order to improve the classification accuracy of unsupervised learning. Supervised learning can get predicated results from a lot of labeled samples which are trained by learning, however the fact of getting the labeled samples needs big support in labor and material. Semi-supervised learning can be trained to get a higher of classification accuracy through a few labeled samples and a large number of unlabeled samples, so semi-supervised learning has been becoming a hot topic of research in AI field. Ensemble learning, integrating a number of classifiers to produce the final classification, obtains a better combined classifier and a better classification performance. As compared with previous single classifier So the combination of ensemble learning and semi-supervised learning can be one of the methods to improve the performance of classifiers with a small number of labeled samples.The integration of ensemble learning based on semi-supervised learning concept is described in this paper, and the advantages of integration of these two learning methods are introduced. Some of unlabeled samples with high confidence and some of unlabeled ones with low confidence selected as well as their label categories are added into the centralized training activities during the whole learning process, and produced a better classification performance. The main work is as follows:(1)To improve the traditional Tri-Training classification algorithm, utilizing three different base classifiers, and employ the weighted accurate rate during the process of the classifier combination to generate the classifier, namely Tri-Training classification algorithm.The traditional one does make use of a classification algorithm to generate three base classifiers whose differences are relatively little. Only through the difference of sample set be its performance would better. However, this paper, adopting three different classification algorithms which would accordingly generate three different base classifiers, is to improve the classification performance according to the different sample sets.(2)A APTTC classification algorithm(Aggregation pheromone metaphor for Tri-Training classification) is proposed, when ant colony aggregation pheromone calculation is integrated into Tri-Training classification algorithm. Ant colony aggregation pheromone concentration is to be calculated and works as the confidence. Unlabeled samples with high confidence(set a threshold, if greater than the threshold, it would be seen as the unlabeled sample with high confidence) are together with their label categories to be mixed into the labeled sets, which would produce different training sample sets and the classification performance meanwhile would be improved.(3)A classification approach based on BTTCR(Bagging and Tri-Training based on Confidence Resampling) is presented by the combination of Bagging and Tri-Training semi-supervised classification algorithm. In each iteration, resampling would be done according to the degree of confidence, and some of high confidence samples and some of low confidence samples as well as their label categories are mixed into the labeled sample sets to be trained. The purpose of selecting high confidence samples is to generate accurate classification and selecting the low confidence samples is to cooperate the training so as to highly different sample sets are produced and the accuracy of classification is improved.This paper is based on COREL image database and the Indoor Scene image set as the experimental data, and the above three of(Tri-Training-3, APTTC and BTTCR) classification algorithms are separately employed to classify images and experiments of comparison are made with the traditional classification algorithm where a kind of classification algorithm is used to obtain three kinds of classifier of Tri-Training(Tri-Training-NB, Tri-Training-KNN and Tri-Training-SVM)classification algorithm. It is proved that the above of three approaches can effectively better and improve the performance of classification algorithm.