The general goal of this project is to develop generative models for images of natural scenes, as well as associated algorithms for unsupervised learning of such models from natural images. The learned models can then be used for image representation and pattern recognition. A particular class of models investigated in this project are the compositional sparse coding models, where the images are represented by automatically learned dictionaries of templates, and each template is a compositional pattern of wavelets that provide sparse coding of the images. The PI and collaborators also investigate related models where the templates are inhomogeneous Markov random fields whose energy functions are defined by the sparse coding wavelets.
Image understanding is at the hearts of many modern technologies. It is also a major function of human brains. The key to image understanding is automatic discovery of patterns in the images. The goal of this project is to develop methods for learning dictionaries of patterns or "visual words" for representing images of our environments. The learned dictionaries can be very useful for image understanding and object recognition.