Image classification using mixed-order structural representation based on mid-level feature

Many successful methods for image classification transform low-level descriptors into mid-level features, and obtain a compact and discriminative representation. However, these approaches rarely think about the spatial context information between mid-level features. In this paper, we present a new mixed-order structural (MOS) representation which may generalize the widely used method spatial pyramid representation (SPR) with correlation modeling method. We first partition the image following SP and extract the mid-level feature for each spatial region. Then our MOS consists of 0^th and 1^st - order correlation model based on mid-level features. The 0^th- order correlation can answer what the mid-level feature represents, and the 1^-th order corresponds to spatial contextual information in ambient neighbors. In addition, we apply principal component analysis to reduce the dimension of mid-level features which makes our MOS can be a relatively low dimensionality. Our experiments demonstrate that the proposed MOS representation achieves a better performance compared with spatial pyramid matching (SPM) and other state-of-the-art algorithms.