A Novel approach to dictionary learning for the bag-of-features model
Siying Zhang, Alex Po Leung · 2017
With the fast development of various methods for image classification using the bag-of-features model, machines can efficiently classify images by image content. Spatial pyramid matching (SPM) for sparse coding to create the dictionary is a popular and very well performing approach for image classification. The linear SPM was proposed to take advantage of the speed of the linear Support Vector Machine (SVM) on spatial-pyramid pooling of SIFT sparse coding. In order to improve k-means clustering in the bag-of-features model, a novel approach to dictionary learning to take advantage of the k-means++ algorithm which provides statistical guarantees for the results of clustering is proposed. Another main contribution is that a novel dictionary learning method is proposed to incorporate more information for the Support Vector Machine (SVM) to build a reliable model, and the information is generated by k-means++ algorithm with different values of k. Our experiments demonstrate that the classification results using our method have been improved a lot with a reduction of computational complexity compared to previous results. In the experimental section, it is shown that our method outperforms the k-means and k-means++ algorithms by nearly 10% and 5% respectively.