Low-resolution Iris Recognition via Knowledge Distillation

Iris recognition with low resolution images is a challenging task in the biometric research community. To achieve a good performance, the common well-established systems require high resolution images, the appropriate camera acquisition settings and the user interaction and cooperation. This raises limitations regarding the applicability of the technology and the level of hygiene in its operation. These requirements are not easy to meet in certain real-world environments like on mobile devices or iris recognition from a distance in border control. To help overcoming these constrains, this work introduces a novel approach for low and extremely low-resolution iris recognition combining Deep Learning and Knowledge Distillation techniques. This work starts by adapting the margin penalty loss to the iris recognition problem. This included novel analyses on the appropriate penalty margin for iris recognition, as the identity nature in iris might differ from the well-studied penalty margins in face recognition. Additionally, this work is the first to build analyses towards finding the optimal deeply learned representation for the identity information embedded in the iris capture. Most importantly, this work proposes a training framework that aims at producing iris representations from extremely low-resolution images in a manner than insures that these representations are similar to those of high resolution. This was realised by the controllable distillation the knowledge of an iris recognition model trained for high-resolution images into a model that is specifically trained for extremely low-resolution irises. The presented approached proved extremely successful and lead to the reduction of the verification errors by more than 3 folds, in comparison to traditionally trained model for low resolution iris recognition.