基于概率超图直推式学习的交互式图像检索方法研究

To bridge the semantic gap that exists between the representation of an image by low-level features and its high-level semantic content as perceived by humans, and to meet user-specific requirements, interactive image retrieval technology has drawn millions of researchers’ interest. However, there are still some problems which have not been solved very well. This project focuses on the key problems of the interactive image retrieval technology. Firstly, for the non-structural content of image, the image is proposed to be modeled by a probabilistic hypergraph. Different kinds of feature points of image are considered as the nodes of the hypergraph. The probabilistic weight of the edges is computed using the information aggregation operator. These constitute an efficient hypergraph representation of image. Secondly, in order to capture the distribution of the image base efficiently and improve the retrieval performance, a probabilistic hypergraph model of the image based is developed. In this hypergraph, one image is taken as a node, and the edge structure is constructed by taking the similarity of images and regional density of the image base into account. This model can provide the constraint conditions of the consequential transductive learning. Thirdly, in aim to exploit the feedback information, this project investigates transductive learning method based on the probabilistic hypergraph. It mainly focuses on the construction of the objective function based on various norms, and the optimization of the solving algorithm. This project can not only enrich the theory of image analysis and recognition but also provide the theoretical basis for the recognition and retrieval of other multimedia.

为了减小图像低层视觉特征和高层语义概念之间的语义鸿沟，同时满足个性用户的偏好，交互式图像检索已成为人们的研究热点，然而它的一些技术问题仍未得到良好的解决，本项目围绕交互式图像检索的若干关键问题展开研究。①针对图像内容的非结构性，提出用概率超图对图像进行建模，综合图像中多种不同意义的特征点形成超图的节点，利用信息集结算子计算超图边的概率权值，形成有效的图像概率超图表示。②为了有效地刻画图像库的分布并提高图像检索效率，研究基于概率超图的图像库分布模型，以图像为节点，同时考虑图像之间的距离和图像库的局部密度建立图像库的概率超图的边集，此概率超图模型可作为后续直推式学习的约束条件。③为了有效地利用反馈信息，研究基于概率超图的直推式学习方法，构建基于不同范数的目标函数，优化其求解方法。本项目的理论研究不仅能丰富图像分析和识别的相关理论，而且能为其它多媒体的识别和检索提供理论依据。

本项目挖掘图像的结构信息，结合图像的低层和中层的特征，对图像建立了多层次和多尺度的图表示，利用信息集结算子，提出了基于加权超图和IOWA算子的图像椒盐噪声滤除算法，提出了基于多层次图和IGOWA算子的图像椒盐噪声滤除方法，同时为了刻画多尺度词袋特征的空间关联信息，提出了基于多尺度属性图核的图像分类，提出了协同关联图的概念并应用于图像集表示，在弱监督的图像语义分割方面，探索了结合最优包围盒和Grabcut的弱监督分割，针对广泛使用的聚类算法，研究了层次聚类算法的评价指标和最优类别数的确定方法，另外，针对目标跟踪，提出了基于拉普拉斯稀疏表示的灰度-热红外环境下的目标跟踪方法，同时研究了复杂环境下基于角点回归的全卷积神经网络的车牌定位，提出了基于卷积神经网络的监控视频人数统计算法。