Using Example-Based Machine Translation Method For Automatic Image Annotation

The paper proposes that the image annotation task can be thought of as similar to the machine translation problem and apply the example-based machine translation method to this problem. The method is based on the idea of performing automatic annotation by imitating annotation examples of images with similar visual scene. It can make full use of both correlation of annotation words and context of image regions in same image. Given an input image, the most visual similar images are retrieved from the annotated images. The annotation words of the retrieved images can be used as the annotation of the input image. From this view, we can say traditional techniques of content-based image retrieval (CBIR) are more apt to the task of automatic image annotation. As an example-based machine translation method, the judgment of visual similarity between images plays an import role. Earth mover´s distance (EMD) is chosen as similarity measure for visual features. In order to make the EMD favor the similar regions between images, an enhanced EMD is presented. The approach does not rely on clustering and consequently does not suffer from the granularity issues. Experiment results show that the proposed mechanism outperforms the state-of-the-art techniques in annotating a large image collection using the same data set and same feature representations