Learning bag-of-words models using sparse partial least squares

Abstract

Representing images using Bag-of-Words (BOW) model has been shown excellent performance for image classification and retrieval. However, there are still some limitations in this model such as the presence of many noisy visual words and the hard to define vocabulary size. To circumvent these drawbacks, this paper concentrates on tuning compact, robust and thus efficient BOW model even with a universal size for image representation. The proposed approach increases expressive power by employing Sparse Partial Least Squares (SPLS) for tuning the traditional and high-dimensional BOW model and learning more discriminative subspace with 10 latent variables. The performance of learning BOW models to image classification is studied through extensive experiments on the VOC 2006 dataset. Empirical results indicate that the proposed method yields quite stable results, and outperforms the traditional BOW models with various vocabulary sizes and PCA with SVM. © 2011 Springer-Verlag Berlin Heidelberg.