Linear regression for Automatic Image Annotation

Abstract

Automatic image annotation is an effective technology to enhance the performance of image retrieval. In order to annotate image accurately, we introduce a novel annotation method based on regression models. Firstly, with different independent views, both the visual and the textual modalities are efficiently represented in a continuous vector space form, and are named by the visual blob vector and the semantic description vector, respectively. Then, instead of mining the association probability model between images and keywords, the task of annotation is reformulated into fitting a rigorous mapping construction between the visual blob vectors and the semantic description vectors using a method based on least squares estimation. Compared with the previous annotation methods, the merits of the proposed method are conceptually simple, computationally efficient, scalable for huge amount of images and do not require any priori knowledge about images and keywords for modeling the task of annotation. With a highly accurate approximation function, the experimental results demonstrate the improvement of annotation performance.