Visual feature coding for image classification integrating dictionary structure

• We propose a feature coding method integrating density structure within the dictionary.
• An approximating solution by finding k-dense neighbors of local feature is developed.
• The proposed coding strategy well preserves both locality and sparsity.
• The proposed coding strategy improves the dictionary with an online learning scheme.

Methods based on the Bag-Of-Words (BoW) model have made a remarkable success in image classification, but many of which do not consider the structure underlying in the dictionary itself in the feature coding procedure. In this paper, we propose a novel visual feature coding strategy by integrating dictionary structure, into which it incorporates not only the relations between visual features and visual words, but the pair-wise relations among visual words. We further develop an approximating solution to the integrated coding procedure by finding the k-dense neighbors (kDN) of a feature, rather than its k-nearest neighbors (kNN). Therefore, the proposed visual feature coding method well preserves similarity, locality and sparsity, and can be used to improve the dictionary itself in turn by proposed specially designed online learning scheme as well. Experiments on three well-recognized benchmark datasets have been conducted and the experimental results demonstrate that it achieves superior image classification performance to other state-of-the-art methods.