Babenko et al. apply convolutional neural networks to image retrieval. In particular, the architecture proposed by Krizhevsky et al. [1] (shown in Figure 1), has been used to compute features for image retrieval. They experimented with different layers, dimensionality reduction techniques (e.g. PCA and Large-Margin Dimensionality Reduction [2]), and both the pre-trained model and a refined model to demonstrate state-of-the-art performance on many image retrieval datasets.

Figure 1 (click to enlarge): Architecture proposed in [1].

• [1] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, 2012.
• [2] K. Simonyan, O. M. Parkhi, A. Vedaldi, and A. Zisserman. Fisher vector faces in the wild. In British Machine Vision Conference, 2013.