Abstract:

The web and image repositories such as Fickr are the largest image databases in the world. There are billions of images on the web, and hundreds of million high-quality images in image repositories. Currently, these images are indexed based on manually-entered tags and individual and group usage patterns. In this work we explore a third information dimension: image features. We explore probabilistic latent semantic analysis (pLSA) in order to infer which visual patterns describe each object. We build models that connect words and image features, and use content features and tags to find similar images. We demonstrate that image features using gray-scale salient points and an aspect model based on pLSA outperforms a conventional word-frequency model as well as refined color-histrogram approach on an image-similarity task.