Statistical Learning Techniques In Image Retrieval

With the development of the network and computer science, imageretrieval is playing an important role in our work and daily life. However,the large gap between low-level features and high-level semantic concepts,and the small-sample-size problem in the learning process of image retrievalprevent this technique from being widely used. On the other hand,statistical learning techniques are a powerful tool for people to understand theworld, whose application in image retrieval is not fully explored. In thisarticle, we consider the problem of applying statistical learning techniques inimage retrieval. Our work can be divided into three parts.Firstly, to solve the small-sample-size problem in image retrieval, wepropose a transductive learning framework named Multiple Random Walk(MRW). From a theoretical point of view, MRW unifies some existingsemi-supervised learning algorithms;from a practical point of view, MRWimproves the performance of image retrieval by a large margin.Secondly, to solve the problem of selecting unlabeled images inrelevance feedback, we propose a new active learning method named MeanVersion Space (MVS). From a theoretical point of view, MVS integratesexisting active learning methods in image retrieval;from a practical point ofview, MVS accelerates the convergence to the query concept, thus improvesthe performance of retrieval systems.Lastly, to solve the problem of pseudo relevance feedback, we propose a novellearning method named Iterative Probabilistic One-Class SVM (IPOCS). From atheoretical point of view, IPOCS generalizes the multi-view algorithms fromsemi-supervised scenario to un-supervised scenario;from a practical point of view,IPOCS makes the full use of the original ranking order provided by web image searchengines, so it can refine their retrieval results considerably.From a theoretical point of view, our work generalizes the semi-supervisedlearning, un-supervised learning and active learning methods, and providestheoretical guide to the future development of image retrieval and related fields;froma practical point of view, our work well solves the small-sample-size problem, imageselection problem in relevance feedback, and pseudo relevance feedback problem.Our work has practical value in transferring image retrieval from experiments to realapplications.