Probabilistic Latent Semantic Analysis Method Based On Dynamic Threshold Model

Object discovery and location is an important branch of computer vision. Withthe explosion of digital images in the internet, the approach of image retrievalbecomes diverse. The global feature based image classification cannot meet all thedemands. Sometimes we need to retrieve an image according to local objectrecognition and matching. Thus, how to discover local objects and their locationsbecomes the key issue of image research area. Probabilistic Latent Semantic Analysis(PLSA) is a topic discovery model used in natural language processing. This modelcan effectively discover the latent structure underlying data. In image processing, ithas successfully used to discover object categories and their approximate spatiallayout. However, PLSA will result in overfitting problems in practice, including thecircumstance of unclear membership of topics given image and the case of highsimilarity between different topics.With this background, this paper explores a dynamic threshold based PLSAmethod. Our work is as the following:(1) With an efficient sparseness constraint ofPLSA, the proposed method will ignore the topics with relatively low membershipsand highlight the impacts of dominating topics in each image. The SparsenessConstraint Threshold is employed to restrain the number of relevant topics within agiven image. This method will effectively filter out the redundant information oftopics and eventually, successfully solve the problem of ambiguous distribution oftopics.(2) Proposed a strategy of mergence based on semantic similarity. For all thediscovered topics, we set the Mergence Constraint Threshold which dynamicallydecides whether to merge the similar topics or not. Our aim is to implement that onetopic can be mapped to only one clear object category, while ensuring the informationof categories completely preserved. This strategy can also build the hierarchicalstructure of topics. hierarchical structure of object will emerge. By modifying theconstraint condition, one can abstract away the topics in different levels. The patternsin different hierarchies can be cognized.The experiment result show that compared to PLSA, dPLSA effectively solvedproblem of overfitting and still possesses the high reliable capacity of latent structurediscovery. Meanwhile, it is successfully extended to the hierarchical model.