Describing and learning of related parts based on latent structural model in big data

Abstract

In this paper, we propose a novel latent structural model for big data image recognition. It addresses the problem that large amount of labeled training samples are needed in traditional structural models. This method first builds an initial structural model by using only one labeled image. After pooling unlabeled samples into the initial model, an incremental learning process is used to find more candidate parts and to update the model. The appearance features of the parts are described by multiple kernel learning method that assembles more information of the parts, such as color, edge, and texture. Therefore, the proposed model considers not only independent components but also their inherent spatial and appearance relationships. Finally, the updated model is applied to recognition tasks. Experiments show that this method is effective in handling big data problems and has achieved better performance than several state-of-the-art methods.