Research On Image Diversity Search Reranking Based On Random Walk Model

With the development of the internet in recent years, the multimedia dataespecially the images and videos are explosively growing. Visual resources have someunmatchable features like intuitive and vivid. Most of people are indispensable tothese resources. Therefore, how to correctly retrieve visual resources is the mostconcern to all search engines. However, mainstream search engines use text retrievalusually based on the title of the image or the description meta-data rather than visualinformation of the image itself, so that the searching results are not ideal. ImageReranking algorithms have been proposed in order to make in line with the need ofthe users. Reranking is generally using the information of the image itself to reorderthe original result in order to improve the user experience.Most of current researches pay more attention to the diversity ReRanking ratherthan Relevance ReRanking. Relevance ReRanking means the relevant results wouldrank in the forefront of the return results, but rarely consider the interrelation of eachother. Diversity ReRanking will significantly improve the diversity of the returnresults also meaning the covering themes to reduce the redundant information.In this thesis, we applied the random walk theory to image ReRanking. Our workincludes:1. The return results can be reordered according to their relevance scores byusing image based relevance reranking algorithm..2.We proposed a Image diversity reranking algorithm based on DivRank.DivRank algorithm introduced Vertex-Reinforced random walk to the reranking fieldin order to reduce the redundancy between the searching results. To this end, weapplied DivRank to the image reranking field and proposed appropriate GDRIDalgorithm.3.We proposed a Image diversity reranking algorithm based on DDRank.DDRank can reduce the side effect brought by DivRank which permanently destroyedthe structure of the originally information network so as to ensure the relevancebesides diversity. To this end, we applied DDRank to the image reranking field andproposed DIR algorithm.