Language Translation and Media Transformation in Cross-Language Image Retrieval

Cross-language image retrieval facilitates the use of text query in one language and image query in one medium to access image collection with text description in another language/medium. The images with annotations are considered as a trans-media parallel corpus. In a media-mapping approach, we transform a query in one medium into a query in another medium by referencing to the aligned trans-media corpus. From the counterpart of results of an initial retrieval, we generate a new query in different medium. In the experiments, we adopted St. Andrews University Library's photographic collection used in ImageCLEF, and explored different models of language translation and media transformation. When both text query and image query are given together, the best MAP of a cross-lingual cross-media model 1L2M (one language translation plus two media transformations) achieve 87.15% and 72.39% of those of mono-lingual image retrieval in the 2004 and the 2005 test sets, respectively. That demonstrates our media transformation is quite useful, and it can compensate for the errors introduced in language translation.