Automatic Image Annotation Using Modified Keywords Transfer Mechanism Base on Image-Keyword Graph

Automatic image annotation is widely considered to be an important yet open problem due to the well-known semantic gap. Recent works show that nearest-neighbor-based annotation approaches are simple and effective. In this paper we use a modified keywords transfer mechanism base on image-keyword unidirectional graph to derive a great annotator. The unidirectional graph describes the relationships between images and keywords, and can be derived from images with human annotations. On that basis, a modified keywords transfer mechanism base on visual neighbors is used to annotate new images. Our method achieves better annotation performance than two of the most advanced annotation methods in terms of precision, recall and F1 metrics on the open benchmark database.