The Reranking Method Based On Image Content

With the rapid development of multimedia technology, a large number of digitalimages swarm into people’s daily lives. The overwhelming of multimedia data makesan urgent requirement to accurately and rapidly find the concerned images and videos.Traditional image retrieval based on text may cause “semantic gap”. Thus, imageretrieval based on content is more and more important and image search rerankingtechnology becomes the hot issue in this area. However, the high-dimension of thevisual features usually causes the problem of “curse of dimensionality” and brings thehigh computational and storage burden. In this paper, we conduct a deep research onnetwork image search reranking, and obtain following achievements:(1)Firstly, we extracted features such as color, edge and histogram feature frominitial retrieval results based on text. Secondly, we improve Locality PreservingProjections (LPP) by incorporating Pearson correlation coefficient and relevancedegree information into it. It can enhance the accuracy of dimension reduction effectfor image features. Finally, we utilize Ranking SVM to train a ranking model andrerank initial results. Dimensionality reduction process could not only find the internalstructure of the data accurately, but also save storage space.(2) Based on the above research, we study further about relevance degreeinformation. We propose a simple reranking method based on pseudo relevancedegrees which utilize the information of relevance degree. With few relevance degreelabels，we calculate the pseudo relevance degrees of other unlabeled images, and thenrerank images based on the pseudo relevance degrees. The proposed image searchreranking method is model-free and does not require iterative so that it has lowcomputational cost and performs effectively.