Research On Reranking Algorithms For Image Search

With the rapid development of storage devices, networks and compression techniques, the number of available media data grow rapidly. How to effectively search such visual data has attracted more and more researchers. Due to the mature development of text retrieval, most search engine are keyword-based for image retrieval. However, the textual information usually contains much noise and mismatch image’s visual information, keyword-based image retrieval results are not satisfying. Based on this, image reranking was proposed. Image reranking refers to improve the initial sort of keyword-based results by image visual features and so on, which makes images of user interested rank front.Image reranking usually means relevancy reranking, image is ranked by relevancy score from high to low. Depending on the use of different framework, image reranking algorithms can be divided into linear combination based, clustering based, classification based, graph based and so on. Depending on the use of different feature, image reranking algorithms can use singlemodal or multimodal features and can use textual or visual features. Most algorithms use multimodal features since the limited describe ability of singlemodal. Using visual feature to improve text retrieval results is reasonable due to the semantic gap between low-level features and high-level semantic, so most algorithms use visual feature. In image reranking algorithms, features are generally assumed independent of each other. A type of feature describes some aspects of an image, and so different features describe different aspects of image which means features are not independent. Based on the above consideration, this paper do two work on image reranking which are introduced as follows:1. Asym-Joint-Rerank. This algorithm considers link among multimodal features, and a similarity matrix is constructed in a graphical modal to solve image reranking problem. There are three kinds of similarity during construction of similarity matrix, including similarity between same modal feature of different images, similarity between different modal feature of one image, similarity between different modal feature of different images. This kinds of similarity matrix effectively simulate the relationship between images.2. Knn-Joint-Rerank. Asym-Joint-Rerank just use a type of visual feature, so image content cannot be fully described and taking into account the consumption of computing resources as the number of images increase. Knn-Joint-Rerank adds three types of visual feature to describe image content and adopts K-nearest neighbor algorithm to improve the calculation process of similarity matrix. This is an effective solution to the two problems in Asym-Joint-Rerank.Finally, experiments are conducted on dataset with several other algorithms to demonstrate the effectiveness of Asym-Joint-Rerank and Knn-Joint-Rerank.