Beyond visual features: A weak semantic image representation using exemplar classifiers for classification

Usually, the low-level representation of images is unsatisfied for image classification due to the well-known semantic gap, and further hinders its application for high-level visual applications. To deal with these problems, in this paper, we propose a simple but effective image representation for image classification, which is denoted as the responses to a set of exemplar image classifiers. Each exemplar classifier corresponding to a training image is learned using SVM algorithm to distinguish the image from others in different classes, and hence exhibits some discriminative information, which can also be regarded as a kind of weak semantic meaning. In such a one-vs-all manner, we can obtain the exemplar classifiers for all training images. We then train a linear classifier with structured sparsity constraints for each image category by taking advantages of the weak semantic image representation. Experiments on several public datasets demonstrate the effectiveness of the proposed method.