Revising Knowledge Discovery for Object Representation with Spatio-Semantic Feature Integration

Abstract

— In large social networks, web objects become increasingly popular. Multimedia object classification and representation is a necessary step of multimedia information retrieval. Indexing and organizing these web objects for the purpose of convenient browsing and search of the objects, and to effectively reveal interesting patterns from the objects. For all these tasks, classifying the web objects into manipulable semantic categories is an essential procedure. One important issue for classification of objects is the representation of images. To perform supervised classification tasks, the knowledge is extracted from unlabeled objects through unsupervised learning. In order to represent the images in a more meaningful and effective way rather than using the basic Bag-of-words (BoW) model, a novel image representation model called Bag-of-visual phrases(BoP) is used. In this model visual words are obtained using hierarchical clustering and visual phrases are generated by vector classifier of visual words. To obtain the Spatio-semantic correlation knowledge the frequently co-occurring pairs are calculated from visual vocabulary. After the successful object representation, the tags, comments, and descriptions of web objects are separated by using most likelihood method. The spatial and semantic differentiation power of image features can be enhanced via this BoP model and likelihood method