基于结构化语义特征的图像检索研究

With the popularity of the social media, there are exist a huge number of images which have weakly supervised labels. How to retrieval these images has become the research hotspot in the field of multimedia. The discrimination and robustness of image features are the basic factors related to the results of image retrieval. Focusing on the weakly supervised situation, different semantic feature embedding methods are designed for distinct weakly labels. Firstly, the semantic dictionary is constructed to obtain the discriminative and robust features based on the correlations between the labels, which are co-occurrence and exclusive. And then, considering there exist high-order correlations between image labels, e.g. visual attributes, a unified framework is proposed to discover the latent discriminative feature representations via embedding these high-order correlations into the original features. Finally, the image captions are used as the weakly labels to achieve these correlations between objects in the images. In this way, the semantic feature representations of complex images are extracted for image retrieval.

随着社交媒体的盛行，互联网上存在着大量的带有弱监督标签的图像。如何快速的检索图像成为多媒体相关方向的研究热点。图像特征的辨别性和鲁棒性是影响图像检索结果的重要因素。在弱监督图像语义信息标注下，本项目依据不同形式的语义标注信息提出了相应的语义及其结构信息的嵌入方案。首先是提出利用构建语义字典的方法，将语义标签及其标签之间的共存性和互斥性嵌入到图像特征表示中以此来提高特征表示的辨别性。考虑到语义信息之间不仅存在二阶的关系，而且存在高阶共存的关系，提出一种统一的算法框架自动挖掘和嵌入高阶语义关系，通过利用该方法能够迅速的将图像中的语义信息之间的高阶结构信息挖掘出来，并且能够自动的获取出来能够提高图像检索结果的语义结构特征。最后，考虑到图像内容中物体之间存在着特定的语义结构，提出利用图像的标题式描述来进行图像中物体之间的语义结构的挖掘，从而获得复杂图像的结构性语语义描述。

随着社交媒体的盛行，用户可以便捷的获取图像并且将其分享到网络中，但是这些图像中既包含正常的图像也有一些内容敏感的图像存在。为能够准确的识别出网络中的图像信息，本项目围绕图像内容理解领域展开相应的理论基础和关键技术的研究，在轮廓图像检索、图像识别、高阶语义关系的挖掘等方面取得了创新性的理论研究成果。共发表学术论文8篇，其中IEEE汇刊论文4篇，CCF A类国际会议2篇。申请国家发明专利3项，省部级技术进步一等奖1项、自然科学二等奖1项。完成基于轮廓图像输入的图像检索和识别原型系统一套，相关研究成果已经在国家多个安全部门开展实际应用。