Document expansion for image retrieval
RIAO '10 Adaptivity, Personalization and Fusion of Heterogeneous Information
| Hi-index | 0.00 |
ImageCLEF2007 photo task is different from those of the previous years in two aspects. The caption field in the image annotations and the narrative field in the text queries are removed, and the example images in the visual queries are also removed from the image collection. In the new definition, the information that can be employed is less than before. Thus matching query words and annotations directly is not feasible. This paper explores the web to expand queries and documents. The experiments show that query expansion improves the performance 16.11%, however, document expansion brings in too much noise and the performance decreases 28.24%. The media mapping method based on an image-text parallel corpus is regarded as query expansion. The results of the formal runs show this method performs the best. Compared with the performance of the models without expansion, the MAP improves about 86.69%~143.12%. Integration of the external and the internal resources gains no benefits in the further experiments.