Iris Segmentation in Challenging Conditions

Iris segmentation is an irreplaceable stage of iris recognition pipeline. Its quality hugely affects overall accuracy. Previously when conditions were mild and controlled the task was solved by image processing techniques and rule based approaches. Nowadays widespread of biometric technologies has relaxed operation conditions for such systems demanding more flexible and robust solutions. Constantly increasing data and sensors availability created fertile field for growth of machine learning methods capable to cope with complex conditions. The latest contributions to iris segmentation were made on this surge by leveraging abundant data and modern machine learning algorithms. In spite of previously achieved great results this work addresses even more challenging conditions that allows iris recognition to be used in wide range of real life cases. Novel CNN architectures are proposed in this work. They were designed to combine the latest achievements in classification and semantic segmentation fields. FCN and SegNet architectures have been picked up as prototypes and were strengthened by residual blocks. This allowed to make lightweight networks that could be shipped on various embedded platforms to successfully operate under less controllable environmental conditions. The approach allowed to obtain 0.93 and 0.92 IoU on original and modified CASIA-Iris-Lamp datasets which is a significant improvement in comparison with the results achieved before.