【摘要】 相关反馈是近年来交互式图像检索领域研究的重要方向。首先提出了基于相关反馈的图像检索系统框架,并在此基础上从机器学习的角度分析了相关反馈学习的算法模型、样本获取、分布密度估计,及其在其特定应用背景下的困难和挑战。进而对图像相关反馈技术的研究现状进行调查总结,从聚类和分类两个方面对各种相关反馈算法在基于内容的图像检索中的应用进行了较为深入地研究和比较。最后对相关反馈技术发展趋势进行了展望,指出了该技术与图像语义抽取、用户模型建立以及软计算技术之间存在的密切关系。更多还原

【Abstract】 Motivated by the fast growth of image databases, content-based image retrieval (CBIR) has receivedwidespread research interest recently, where a typical user query is represented by a dynamic combination ofvisual and semantic descriptions of the desired image or class of images. Unfortunately, users often have difficultyspecifying such descriptions. To alleviate those problems, users’ queries are often refined interactively throughmining the relevance of the feedback images, with the parameters combining the features automatically adjustedto adapt to the users’ original needs. The aim of this paper is to clarify some of the issues raised by this new technology by reviewing the currentcapabilities and limitations of the relevance feedback (RF) techniques from the viewpoint of machine learning. A concept CBIR framework with RF techniques embed in is proposed first to facilitate the description, and the basiccomponents of a RF-based CBIR system are illustrated as well. By using the proposed model, the paper proceeds with reviewing various RE techniques existing in theliteratures from three aspects. The data which can be used as the source of RF-mining is investigated firstly. Toanalysis the obtained feedbacks, some assumptions about the distribution of the data must be made. We describetwo kinds of probable distribution assumptions commonly seen in literatures, i. e., Gaussian distribution andmixture models, and the advantages and shortcomings of both assumptions are compared. Thirdly, we identifysix challenges one may encounter in the practice of applying RE technique in CBIR context and point out thatthere is not such perfect algorithm that can deal with all the mentioned difficulties well at the same time. Next, we focus on the classical RE methods, which mainly originate from the pattern recognition or machinelearning area. Roughly saying, we survey the typical RF algorithms from two branches, i.e. the clustering-basedand classifying-based algorithms. The main difference between the two kinds of algorithms lies in that theclustering-based algorithms care more about the understanding of the "query point", which is usually consideredas the semantic representation of user’s query in his or her mind, while the classifying-based ones try to find theclass boundary directly from the image database using some prior knowledge about the statistical structures of thefeedback data. However, it is worth mentioning that the boundary between the two branches of algorithms is softin nature. The merits and shortcomings of the classical RF methods are also discussed. Finally, several typical CBIR systems, either commercial or academic, with RF-techniques embed in aresurveyed The surveyed systems include such famous systems as QBIC, FourEyes, PicHunter, etc. Therelevance feedback techniques adopted in those systems are identified and emphasized. In summary, we think that human is a key factor in the running of the whole CBIR system and it wouldplay a more important role in the operation of the system than now. RF techniques, as a way to incorporate manin loop, will continue to receive wide interests from researchers. The paper then suggests several future promisingresearch directions through analyzing the close relationship between relevance feedback technique and theabstracting of image semantic, user modeling and soft computing.更多还原