In these days there is a large number of deployments for biometric systems. One of the biggest deployments is the India Aadhaar Identification program with more than 1 billion registered citizens. In addition there are many border control programs that use biometric systems for identification. Hence one can suggest that biometric systems become an integral part of the digital world. Iris is one of the most frequently used biometric characteristic in operational systems because the recording is convenient and these systems have a high reliability. Due to the increasing size of the operational deployments of biometric systems, the requirements in terms of, among others, biometric performance, efficiency and reliability increase. Especially the large number of glasses wearers is a challenge for iris recognition systems.
There exist some studies which showed that glasses can deteriorate the biometric performance of iris recognition systems, but none of these showed the causes of this deterioration in detail. We analyzed the influence of glasses on the performance of an iris recognition system with different experiments and we discovered, that the biometric performance loss of iris recognition systems correlates strongly with the subjects wearing glasses. A possible solution to solve this problem is to automatically detect glasses and handle such attempts separately. We propose 3 approaches to automatic detection of glasses and perform a comparative assessment of their accuracy. The proposed approaches are based on: an explicit algorithmic approach with edge and reflection detection, a deep learning approach and an approach using filters of binarized statistical image features. The benchmark that we used was carried out on the CASIA-IrisV4-Thousand database, which contains 20000 near-infrared eye images; 5336 with and 14664 without glasses. The explicit algorithm achieved a classification accuracy of 97.18 %, the statistical approach achieved an accuracy of 98.08 % and the deep learning approach achieved an classification accuracy of 98.97 %. When using a fusion of all three approaches we were able to classify 99.06 % of images with glasses and 99.71 % of images without glasses. Thereby we were able to classify 99.54 % of all images on the CASIA-Thousand database in glasses and non glasses images. With rejecting the detected glasses on the CASIA-Thousand database, we increased the iris recognition performance from 9.19 % to 6.61 % EER.