Saliency Aware Locality-preserving Coding for Image Classification

Quan Fang, Jitao Sang, Changsheng Xu

Abstract: The Bag-of-Features (BOF) model is widely used for image classification. Most BOF models incorporate a step of maximum pooling to generate the raw image representation, where salient atoms with maximum response are reserved for final representation. However, recent locality-preserving coding schemes do not account for the saliency characteristic during the process of generating the raw image representations. In this paper, we propose a saliency aware locality-preserving coding scheme by explicitly considering saliency into the dictionary creation and feature coding stages. The novel coding scheme guarantees strong response in the pooling operation and thus contributes to a discriminative image representation. Experiments on three benchmark datasets validate the effectiveness of the proposed method.