Group encoding of local features in image classification

Saliency is an important factor in feature coding, based on which saliency coding (SaC) has been proposed for image classification recently. SaC is both effective and efficient in case of a moderate-scale codebook. However, empirical studies show that SaC will lose its superiority as the codebook size increases. To address this problem, we propose a group coding strategy, wherein the latent structure information of a codebook is explored by grouping neighboring codewords into a group-code. We apply group coding to SaC and derive the group saliency coding (GSC) scheme. Thorough experiments on different datasets show that GSC consistently performs better than SaC, and also outperforms other popular coding schemes, e.g., local-constrained linear coding, in terms of both accuracy and speed.