A Comparative Study of Global and Local Feature Representations in Image Database Categorization

Content-based image retrieval systems can automatically extract visual content of images which allow users to query images by their low-level features (such as color and texture). However, users usually prefer querying images based on high-level concepts such as keywords. Classifying images into a number of categories (or image classification) facilitates search in image databases. However, the classification performance is heavily dependent on the use of features. In general, there are three feature representation methods, which are global, block-based, and region-based features. As related work only considers using one of these three methods, this paper aims at comparing each of these methods and their combinations by using a standard classifier (i.e. k-nearest neighbor) over thirty categories. The experimental results show that the combined global and block-based feature representation performs the best. In addition, larger numbers of training examples produce higher classification accuracy.