Improving automatic image annotation: Approach by Bag-Of- Key Point

Automatic image annotation is to associate each image a set of keywords and describing the visual content of the image using an automatic system without any human intervention, many approaches have been proposed for the realization of such a system However, it is still inefficient in terms of semantic description of the image. Recent works show a frequent use of a special technique known as bag-of-key points that describes an image as a set of local descriptors using a histogram. Each bin of the histogram represents the importance of a visual pattern (called visual word) in the image. But crucial representation choices such as the choice of local features, the steps of building the visual vocabulary have not been thoroughly studied in existing works. In this paper, a novel approach based on Scale Invariant Features Transform (SIFT) features and treatment of the different steps of building de vocabulary are proposed. The proposed approach creates more robust signatures for images and better reflects the weight of visual words. The categorization of images has been the subject of the second phase of this approach. The purpose of this phase was to find a classification model that best suits the index method proposed, while avoiding problems due to large data and large dimension. Experiments with Corel-1000 dataset demonstrate that the proposed improvements outperform known techniques in scene categorization.