基于语义区域提取的图像重排

针对目前图像搜索引擎难以正确把握用户真正意图的问题, 从爬虫Web图像搜索引擎检索结果入手,提出三种聚类算法来提取海量Web图像中的语义区域. 这三种聚类算法包括确定初始化中心的K-means聚类、确定参数的最大期望聚类以及基于半监督的K-means聚类算法. 然后选取显著值较大的显著区域作为语义区域.实验分析比较了三种聚类算法的有效性, 最终实现的图像重排系统能比网络搜索引擎更好地反馈给用户精确而且有序的查询结果.

Image Re-ranking Based on Extraction of Semantic Regions

It is difficult for current image search engines to accurately grasp the real intention of users. Based on the search results, we propose three clustering algorithms to extract semantic regions of Web images. These methods include K-means clustering with determined k centers, expectation maximization clustering with the determined parameters, and semi-supervised K-means clustering. We then select the salient regions with the high salient scores as the semantic regions. We demonstrate the experimental results by comparing the three clustering algorithms. The proposed image re-ranking system can more accurately show the ordered search results than web image engines.