Object Recognition Based On Shape And Semantic Modeling

This paper studies the object recognition based on image grammar in shape space. The And-Or graph stochastic grammar is used for the understanding, decomposing, learning and sampling on shape object.A layered graph match algorithm is proposed for the shape matching problem. This method integrates the graph partitioning and matching with the sampling in a Bayesian framework to guarantee the correctness, and can be used for top-down verification in recognition.An explicit local shape descriptor are defined as Graphlet, together with its detecting and learning method. According to the sparse coding theory, 14 single graphlets and 20 composite graplet are learned as codebook for bottom-up discrimination.Based on the And-Or graph model, a cascade framework combining the bottom-up discrimination and top-down verification is implemented. Firstly, by the learned And-or graph of different categories of data, generative templates are sampled under the SCFG/MRF, which contain new instances that have not appeared in annotated training set, that can represent current object category manifold more general. Secondly, the distributions on object categories of 34 graphlets in the codebook are learned by the training set of generative templates. Thirdly, based on the graphlet distributions, the bottom-up cascade pruning can extinct the false candidates rapidly while keeping the true positive. Finally, the rest candidates are send to the graph match verification for final recognition. Experiments show that such framework can achieve satisfied result.In the last part of this paper, an implementation of database for large scale annotated dataset together with the And-Or graph image grammar is proposed. It is the first database that deals with high level visual knowledge in current public large scale dataset.