Semantic middleware: multi-layer abstract semantics inference for object categorization

In this paper, we present a hierarchical model, named as Multi-layer Abstract Semantics Inference (MASI), based on Bag-of-Visual-Words (BoVW) to solve the problem of universal image categorization, including typical and zero-shot image categorization. An abstract hierarchical semantics learning method is proposed in the training step by extracting and selecting abstract visual words in a bottom-up way to train abstract semantic classifiers. For a testing image, its category is estimated layer-by-layer from top to bottom according to its corresponding hierarchical categories. Experimental results on popular image datasets have shown that the proposed method achieves better performance compared with traditional learning methods.