Combining Semantic And Visual Information In Image Reranking With Long Queries

Witnessing the booming of social photos sharing on the Internet, images has become an essential resource in our daily life. In order to improve the users’ satisfaction with the results of image retrieval, visual reranking techniques have been proposed. But for long queries, they are more complex than simple keywords-query as they express highly specific information needs. Thus, in spite of the great success for simple keywords-queries, current social image retrieval do not perform well with long queries. The existing approaches generally utilize visual features to improve the retrieval performance, however, on image understanding, it is usually achieved by semantic information. Based on that, we propose a new image reranking algorithm for long queries considering both the semantic contents and visual features. Specifically, the semantic relevance is established on the multimedia corpus as well as the WordNet and WikiPedia, then perform weighted sum of the underlying visual relevance scores, a new ranking list is obtained. In this paper, we focuses on researching long query image reranking method, the main work and innovations are as follows:1. we regard the long query as multiple visual concepts. The search results of visual concepts can reflect some characteristics of the long query, and it has high accuracy rate, and then, it can enhance the accuracy of the correlation estimates between the long query and the image.2. It proposes a probabilistic model to analyze the relevance of visual concept and initial search results and long query. This method tends not to be affected too much by the initial ordering of images, it overcome the defects of the traditional image reranking method.3. It propose a new image reranking algorithm for long queries considering both the semantic contents and visual features, To estimate the relevance score, we linearly combine these two measures. Then a new ranking list is obtained. Among them, the semantic relevance is established on the multimedia corpus as well as the WordNet and WikiPediafor they are more accord with human perception, The experiments results demonstrate that the proposed method can effectively improve the performance of image reranking.