A comparative image auto-annotation

Abstract

The image annotation is an effective technology for improving the Web image retrieval. Many works have been proposed to increase the image auto-annotation performance. The annotation quality depends on many factors: segmentation, choice of visual features...etc.. An image contains several visuals information (color, texture, shape….) but the best choice of these parameters is a difficult problem. The main focus of this paper is two-fold. First, we compare between image annotations in latent space and textual space. Second, we survey influence of visual features choice on image annotation. For that, we developed two image auto-annotation systems (AIA-LSA, AIA-WLSA). For each one, we propose three prototypes based on different visual descriptors; such as: texture, color and fusion of both parameters.Corel data set is used to experiment our prototypes, the results show that the annotation by latent space is more efficient than the annotation by textual space and the best choice of visual features is a persistent problem in general images.