Active Learning-based Multi-label Image Classification

With the rapid development of Internet information technology, the quantity of image data is exploding. In the face of massive image data, how to effectively manage the vast amounts of image data has become an urgent problem to be solved. Image classification is an effective way to manage the image data. Most early image classification methods focus on the binary or multiclass classification problem, only need to select a category label for image every time. It is the most common single label classification method. These single label learning methods are simple and feasible, and have achieved great success in practical application. However, in the real world, an image often contains more than a single semantic, therefore the multi-label learning method has more significance in practical application.In order to reduce the labeling cost of obtaining the image data, the thesis focuses on training sets selection technology in multi-label image classification, and the research scheme of active learning-based multi-label image classification method is proposed. The whole research work is as follows:(1) Research the sampling strategies of multi-label image classification based on active learning algorithm. Research and analysis the existing two kinds of methods: query based on sample and query based on sample label pair. Analysis research ideas and superior or inferior of typical algorithms, it provides a theoretical basis for subsequent research.(2) For the problem that redundancy of label information may exist among the method which query instances only, take the uncertainty sampling strategy into consideration, and propose the optimal label subset selection method. Improve multi-label classification performance and greatly reduce the labeling cost, enhancing the efficiency of classification method.(3) To discuss the problem of mining and utilization of the label correlations in multi-label learning, propose the label correlations mining method based on the hierarchical label tree. Make full use of data information in the sampling process, combining with the inherent label correlations, to further improve the performance of classifiers.In this thesis, the proposed method is demonstrated. Select four evaluation criteria and compared with the existing methods in multiple image data sets. The experimental result is analyzed to validate accuracy and effectiveness of the proposed algorithms.