Asymmetric learning using a cascade codebook for image classification

With huge amount of visual data being online, the need for automatic illegal images filtering system becomes increasingly urgent. A key ingredient in the design of such systems is the asymmetric image classification module. Among the proposed strategies, bag-of-features models together with kernel based classifiers have demonstrated impressive performance. However, when dealing with this asymmetric learning problem, their efficiency is often impaired. This may own to the fact that, a universal codebook is not adequate enough to deal with this kind of problem, which make the classification task difficult. This article proposes a novel approach, where a cascade codebook with a sequence of node codebooks is used to represent images. The experimental results show that, this novel strategy outperforms the approaches which use only one universal codebook.