Effective image representation based on bi-layer visual codebook

Recently, the Bag-of-visual Words (BoW) based image representation has drawn much attention in image categorization and retrieval applications. It is known that the visual codebook construction and the related quantization methods play the important roles in BoW model. Traditionally, visual codebook is generated by clustering local features into groups, and the original feature is hard quantized to its nearest centers. It is known that the quantization error may degrade the effectiveness of the BoW representation. To address this problem, several soft quantization based methods have been proposed in literature. However, the effectiveness of these methods is still unsatisfactory. In this paper, we propose a novel and effective image representation method based on a bi-layer codebook. In this method, we first construct the bi-layer codebook to explicitly reduce the quantization error. And then, inspired by the locality-constrained linear coding method^[18], we propose a ridge regression based quantization to assign multiple visual words to the local feature. Furthermore, the k nearest neighbor strategy is integrated to improve the efficiency of quantization. To evaluate the proposed image representation, we compare it with the existing image representations on two benchmark datasets in the image classification experiments. The experimental results demonstrate the superiority over the state-of-the-art techniques.