搜索引擎中基于内容的图像重排序

Content-based image re-ranking technology in search engine

针对基于文本的图像搜索结果的排序不能很好满足用户查询期望的问题,提出两种基于内容的图像搜索结果重排序方法:基于相似性积分的重排序算法(SI算法)和基于Dijkstra算法的重排序算法(D算法)。这两种方法把图像作为节点,利用图像的颜色和形状特征计算图像间的相似性,并将相似性作为边的权重构建相似性图,SI算法根据每个节点图像相似性积分的大小来进行排序,D算法则利用Dijkstra算法从指定图像开始对所有图像节点进行遍历排序。实验结果表明,两种方法都能很好地改善图像搜索排序结果,其中SI算法适合使用在初始查准率在0.5～0.9的情况;而D算法不要求初始查准率,但对图像间相似性值的精确度要求高,可以用于用户指定一张查询相关图像的重排序。

As the existing text-based image search results sorting cannot meet the users＇ query expectations, two kinds of content-based re-ranking methods for image search results named SI （ Similarity Integral） algorithm and D （Dijkstra） algorithm were put forward. These methods treated images as nodes, used the color and shape features to calculate the similarity between images, and took the similarity as the edge＇s weight to construct the similarity graph. SI algorithm sorted the images according to the similarity integral of each node image, and D algorithm traversed all the images from the specified image by Dijkstra algorithm. The experimental results show that both of the methods can improve the sorting performance of the image search. In addition, SI algorithm is suitable for the situation with initial precision rate at 0.5 - 0.9, while D algorithm does not require the initial precision rate, but has high accuracy requirements of similarity value between images, and can be used to the images re-ranking queried by an specified image.