Compact image representation based on variability analysis

Image representation is the most fundamental and important aspect in image classification tasks. Most existing image representation methods use quite high dimensional feature vectors for image representation in order to achieve desired performance, which results in an inevitable drawback which is a classification problem with very high-dimensional feature vectors. Meanwhile, the existing methods have not considered image variations in image representation. Thus, an image representation method was proposed to model the variability in image classification. First, a Gaussian mixture model (GMM) was used to model the low-level visual feature vectors. Then, the sufficient statistics of images were constructed. Finally, the proposed variability analysis was utilized to decompose the sufficient statistics, and a compact image representation was obtained by means of partial least square regression. The proposed method not only achieves better performance on the public image classification datasets, but also reduces the burdens of classifier training and feature storage.