流形上认识的特征提取及多特征排序的网络图像检索研究

With the development of Internet and multimedia technology, the demanding for network image retrieval has been greatly promoted in many research fields. However, due to the complex of image content, it is difficult to retrieve related images with high efficiency and effectiveness. The different understandings for data images in the fields of interdisciplinary research, such as brain science, cognitive psychology and information science, provide new ideas for the development of network image retrieval. And thus it is significant and practical to integrate cognitive psychology and information science to further improve the accuracy and efficiency for image retrieval. At present, there are some main problems in image representation and retrieval: descriptors do not have great perceptual uniform, the discriminative projection of local descriptors is weakly robust, and multiple features are limitedly compatible with graph-based ranking. This project aims to solve these problems and then starts the research about network images. The detailed projects are illustrated in three aspects: 1) integrating perceptual uniform analysis and gestalt psychology to construct image representation; 2) the semi-supervised dimension reduction of local descriptor based on topology and gestalt psychology, and the update of visional words; 3) the graph-based weighted fusion of multiple features based on cognitive psychology and image ranking based on the profit of each class. The research works are extending cognitive science to the field of image retrieval, and meanwhile lay the foundation for the further development of network data images.

随着数字、移动互联网及多媒体的飞速发展，极大的推动了各领域对网络图像搜索的需求。但网络图像内容复杂，增大了检索的难度。脑科学、认知科学与信息科学交叉的图像理解为网络图像检索提供了新的思路，如何利用信息科学与认知科学的融合提升网络图像检索的精度和效率是十分重要而有意义的工作。本项目针对目前网络图像特征提取视觉一致性不高、局部描述子判别投影鲁棒性不强以及多特征图排序中的问题，开展网络图像的研究。具体拟研究：1）视觉一致技术及格式塔心理学的图像特征提取；2）拓扑学及格式塔心理学的局部描述子的半监督降维及视觉词汇更新；3）认知心理学的多特征联想图权值融合及分组收益的图像排序。项目研究成果将拓宽认知科学在图像检索领域的应用，同时为网络图像检索的进一步研究奠定基础。

随着数字、移动互联网及多媒体的飞速发展，极大的推动了各领域对网络图像搜索的需求。但网络图像内容复杂，增大了检索的难度。脑科学、认知科学与信息科学交叉的图像理解为网络图像检索提供了新的思路，如何利用信息科学与认知科学的融合提升网络图像检索的精度和效率是十分重要而有意义的工作。本项目针对目前网络图像特征提取视觉一致性不高、局部描述子判别投影鲁棒性不强以及多特征图排序中的问题，开展网络图像的研究。针对视觉一致技术及格式塔心理学的图像特征提取，本项目采用颜色-基元双视觉一致法，提取出了一种新型的感知一致描述子。针对局部描述子问题，本项目提出了一种拓扑学及格式塔心理学的局部描述子的半监督降维及视觉词汇更新方法，最后，本项目实现了认知心理学的多特征联想图权值融合及分组收益的图像排序。本项目共发表SCI学术论文9篇。该研究成果将拓宽认知科学在图像检索领域的应用，同时为网络图像检索的进一步研究奠定基础。