Research Of Interactive Image Retrieval Based On Semi-supervised And Ensemble Learning

With the rapid development of multimedia technology and the increasing popularity of Internet technology, the types and quantities of digital image become more and more, and how to search the requirements of users from the huge quantities of image data quickly and accurately has received more and more concern from by academic and business. In the context, content-based image retrieval (CBIR) technology emerged, which indexes images by the low-level visual features and retrieves images by comparing the similarity between features with more objectivity. Unfortunately, the technology exist the inconsistency between visual features of the low-level and semantic concept of the high-level, so called "semantic gap" problem. On this station, the relevance feedback (RF) technology was introduced, and formed an interactive technology. Up to now, lots of relevance feedback algorithms have been proposed, in which support vector machines (SVM) based relevance feedback methods have received more concern. However, most SVM-based relevance feedback methods are challenged by the small example problem and the asymmetrical training examples problem. In order to deal with the problems effectively, in this paper we propose new schemes as follows to improve the existing SVM-based relevance feedback methods:1. To cope with the small example problem and the asymmetrical training examples problem, an asymmetric semi-supervised SVM (AS3VM) scheme is presented in this thesis. AS3VM focuses on processing the positive and negative examples with different learning strategies. Concretely, in each round of RF, QPM is used to generate a virtual positive example; meanwhile, a random subset of the unlabeled images after data cleaning is regarded as the additional negative examples. Moreover, the unlabeled images with lowest confidence judged by the classifier are prepared for the user to label, i.e. active learning relevance feedback.2. To further improve RF performance, by introducing the random subspace theory into AS3VM-AL, an improved RF scheme named RS-A3SVM-AL is proposed. RS-A3SVM-AL aims at training multi classifiers in different random feature subspace, and combines them by using Bagging technique, with the goal of highlighting the feature subspaces which are helpful to enhance current concept searching. The experimental evaluations show that the proposed asymmetric semi-supervised learning methodology in conjunction with random subspace theory is effective to improve CBIR performance and achieves better performance than some existing approaches.