数字图书馆基于集成学习的相关反馈图像检索系统研究

Research on Image Retrieval System of Relevance Feedback Based on Integrated Learning in Digital Library

为充分利用数字图书馆中图像资料所蕴含的各种信息并对其图像检索系统进行合理优化,在全面提取图像各种特征的基础上,引入机器学习的集成学习技术,并将其中的加权投票法和意见一致性方法应用在数字图书馆反馈图像检索中。首先提取物理层特征包括:图像的颜色、纹理、轮廓形状等信息以及逻辑层特征包括:图中对象间的空间关系特征以及抽象层特征包括:场景、情感、行为特征,形成各子特征集。在各子特征数据集上通过欧式距离法进行相似性度量,以产生差异性的子检索结果,并通过集成学习的加权投票法和意见一致性方法将各子检索结果进行加权投票,这也是个决策级结果融合的过程。权重根据人机交互时用户的相关反馈进行动态调整,反复循环,直到用户得到满意的检索结果。该系统充分发挥图像中包含的不同特征的优势,在此基础上构建更加合理的图像检索系统。该系统可以更好地满足用户在数字图书馆进行图像信息检索时多样化、个性化用户要求。

layer features include:scene,emotion,behavior features,forming each sub feature set.In order to produce different sub retrieval results,we use Euclidean distance method to measure the similarity of each sub feature data set,and use the weighted voting method of integrated learning and consensus method to weighted vote the results of each sub search,which is also a process of decision-making level results fusion.The weight is dynamically adjusted according to the user's relevant feedback during human-computer interaction,and it circulates repeatedly until the user gets a satisfactory retrieval result.The system gives full play to the advantages of different features contained in the image,and on this basis,builds a more reasonable image retrieval system.The system can better meet the requirements of users'diversification and personalization in image information retrieval of digital library.