The Dual Codebook:Combining Bags of Visual Words in Image Classification

In this paper, we evaluate the performance of two conventional bag of words approaches, using two basic local feature descriptors, to perform image classification. These approaches are compared to a novel design which combines two bags of visual words, using two different feature descriptors. The system extends earlier work wherein a bag of visual words approach with an L2 support vector machine classifier outperforms several alternatives. The descriptors we test are raw pixel intensities and the Histogram of Oriented Gradients. Using a novel Primal Support Vector Machine as a classifier, we perform image classification on the CIFAR-10 and MNIST datasets. Results show that the dual codebook implementation successfully utilizes the potential contributive information encapsulated by an alternative feature descriptor and increases performance, improving classification by 5-18% on CIFAR-10, and 0.22-1.03% for MNIST compared to the simple bag of words approaches