Hierarchical density estimation for image classification

Histogram (bag-of-words) and Gaussian mixture models (GMMs) have been widely used in patch-based image classification problems. Despite the satisfactory results reported, both methods suffer from a number of disadvantages. For instance, a histogram may be easy to learn but has a large quantization error; on the contrary, Gaussian mixture model based methods have better modeling capabilities but are inefficient in both learning and testing. In this thesis, we present a novel hierarchical density estimation approach for image classification. This new approach partitions the feature space into small regions using a tree structure. For each region, "local" distribution is characterized by class-conditional Gaussians via hierarchical maximum a posteriori (MAP) estimation. We further enhance the parameter estimation by smoothing over a collection of randomized trees. This new approach enjoys the merits of superior modeling capability, robust parameter estimation, and efficient testing. Experiments on scene classification demonstrate both the effectiveness and efficiency of this new approach