The Application Of Active Learning Method Based On Multi-features In Image Classification

With the development of computer, network, communication technology, various forms of digital image have been growing explosively. In the image classification management, image classification research has important application significance. In image classification, a lot of labeled samples are needed, while marking sample takes a lot of manpower and material resources. Therefore, active learning method, for reducing the label cost, has received the widespread attention in the study of image classification.The basic idea of active learning method is to select samples with larger amount of information to train classification model. So that less marked samples could be used to achieve expected classification accuracy. The data in the process of active learning, however, are usually based on single feature, and the amount of information selected tagging samples is limited. To further improve the label sample information, this paper studied the active learning based on multiple characteristics, and it is adopted into image classification. In the process of active learning method based on multiple characteristics, the paper mainly studies the feature selection methods, and the sampling strategy in active learning method. In feature selection method study, a number of SVM based on different characters are integrated to solve the construction problem of the classification model based on multiple characteristics at first. And then by analyzing the multiple features selection method, an improved EPD feature selection method is proposed. The previous sampling strategy is based on a single sample feature, so it is not suitable to be directly used in active learning method based on multiple characteristics. In order to realize the sampling based on samples with multiple features, this paper puts forward an improved edge sampling method. It forecast unlabeled samples on various SVM models, and through forecasting result and distance from the sample to hyperplane, it can select uncertainty sample. In the paper, it is called "multiple marginal sampling based on various SVM".Samples in data set of Corel and NUS-WIDE-CATEGORY are used to verify this method, and to carry on comparative analysis with other methods. Experimental results show the proposed method can effectively improve the image classification accuracy.