Adaptive appearance modeling via hierarchical entropy analysis over multi-type features

The descriptiveness of visual models is crucial for many image processing applications, however, it is still challenging to adaptively formulate such models. This paper systematically advocates a generic and adaptive appearance modeling method. For object-specific instances in images, it can adaptively generate a descriptive codebook by exploring the maximum discriminability of multi-type features. The key idea is to define feature-independent information entropy as a unified criterion to measure different features in a common entropy space. Towards this goal, a hierarchical maximum entropy (HME) model is proposed to conduct multi-feature selection based on the random forest. Specifically, the improved random forest replaces space-specific expression “distance similarity” with the statistical concept “entropy”. Thus, the random forest could integrate the subspace clustering results from different feature spaces. Such integration can not only afford adaptive feature selection and cross-feature error control but also be robust to possible weak/inconsistent feature expressions. To effectively construct a class-specific appearance model, a sparse codebook model, consisting of a series of weak learners, is proposed to further explore the maximum discriminative subspaces of each object class. Finally, a maximum entropy model is proposed to formulate appearance model by optimizing the probabilistic distributions of all the codebook words’ response parameters. To verify the efficacy and effectiveness of the proposed model, it is applied to multi-class image classification. We conduct extensive experiments and make comprehensive evaluations w.r.t several state-of-the-art methods over PASCAL VOC 2007, VOC 2012, Caltech 101 and Caltech 256 datasets. All the results demonstrate the advantages of the our method in terms of precision, robustness, flexibility, and versatility.