The Research Of Image Classification On Combining Semi-supervised Learning And Active Learning

With the rapid development of multimedia technology and the universal of the Internet, the number of digital images shows the rapid growth. It is the focus that how to retrieve, classify, excavate, and use these information. In particular, the emergence of cloud brings greater challenges for quick retrieval. Artificial data analysis often consumes too much time, and has not kept the pace with the requirement for data generation and use limitation. In this context, machine learning receives more and more attention, as an important method for automated data analysis.The main idea of machine learning is that uses computer to simulate human learning activities. It is a research technique that computer recognizes existing knowledge, acquires new knowledge, and continuously improve the performance. According to forms of learning, machine learning can be divided into supervised learning and unsupervised learning, while supervised learning can be divided into full-supervised learning, semi-supervised learning and active learning. The classification accuracy of supervised learning is not high, and supervised learning requires sufficient training samples, while labeling samples is very time consuming. But semi-supervised learning and active learning both use labeled instances and unlabeled instances. It can maximize the performance of the classifier with small amount of labeled instances, that combining active learning and semi-supervised learning to the image classification.This article combines active learning based on the semi-supervised learning. In the samples selection process of the semi-supervised learning, not only these samples with highest confidence are added to the training set, but also the most controversial example is selected to expert system to label, and then the example is added to the training set. The main work in this paper is as follows:1) Classic semi-supervised algorithm---co-training algorithm requires the sample set with sufficient redundancy view. That requires that samples have two different views, can able to learn a strong classifier. But in practice, the requirement is difficult to meet. In this paper, Bagging algorithm and RSM algorithm are used to divide the initial labeled data set into two2) In the whole learning process, two classifiers are needed that one is trained by labeled data, while the other classifier is trained by labeled data and some unlabeled data. In the paper, since the component of training set for two classifiers are different, the two model can describe the unlabeled data set from different aspects.3) In this paper, combining the advantages of semi-supervised learning and active learning not only exploits the information of the abundant unlabeled samples, but also can reduce the rate of adding sample labeling error to training set. Experiments show that the method achieves comparable results compared with some traditional approaches.