Deep Sparse Feature Selection and Fusion for Textured Contact Lens Detection

Distinguishing between images of irises wearing textured lenses versus those wearing transparent lenses or no lenses is a challenging problem due to the subtle and fine-grained visual differences. Our approach builds upon existing hand-crafted image features and neural network architectures by optimally selecting and combining the most useful set of features into a single model. We build multiple, parallel subnetworks corresponding to the various feature descriptors and learn the best subset of features through group sparsity. We avoid overfitting such a wide and deep model through a selective transfer learning technique and a novel group Dropout regularization strategy. This model achieves roughly a four times increase in performance over the state-of-the-art on three benchmark textured lens datasets and equals the near-perfect state-of-the-art accuracy on two others. Furthermore, the generic nature of the architecture allows it to be extended to other image features, forms of spoofing attacks, or problem domains.