Effects of Modifying Features While Applying Co-occurrence Model on Blob Vocabulary for Automatic Image Annotation

In order to provide multimedia (e.g. image/video) retrieval capabilities, digital libraries have depended on manual annotation of images. Providing a way to perform annotation automatically would be more useful, efficient and scalable with expanding image collections. In this paper, we study the Co-Occurrence Model [2] where, by training on a set of images with given keywords, we can automatically annotate test images. The model assumes that each image may be described using a small vocabulary of blobs. The blobs are created by segmenting images into regions and then clustering features computed over each region. The model is based on calculating co-occurrence probabilities of words and blobs. It was experimentally observed that RGB color model features extracted from the images provide better annotation performance than Lab color model features. Moreover, by modifying the weights of various features and normalizing them, we can increase the number of retrieved words with improved average recall and precision over all queries using the model.