Smart Iris Classification Using Weighted Average Ensemble Learning

Abstract

Developing a higher-level security system for iden- tification or authentication has long been an active research subject in many fields. Traditional security systems utilize a key or a password to secure a process or a product, whereas bio metric security systems use a person’s physical or behavioral attributes. Iris patterns play a significant role in a number of potential recognition or authentication applications because of their uniqueness, universality, dependability, and stability. The use of iris recognition techniques in bio metric identification and authentication systems has increased significantly. In this work, a novel approach for classification of iris is presented, making it simple for anybody to apply this technology. This model allows for the usage of any eye image and only selects photos that pass the model’s internal filters. Besides, this study provides iris identification model that starts with eye detection and ends with iris image recognition. In addition, a method for iris classification is presented in this study that combines Transfer learning and Convolutional Neural Networks (CNNs) algorithms using Ensemble learning. The automatic segmentation technique for iris detection uses Hough Transform and is capable of localizing the pupil and iris region, as well as obstructing eyelids, eyelashes, and reflections. To overcome the image irregularities, the iris region is extracted and then extracted iris is converted into rectangular block using Normalization. In this paper, a weighted ensemble technique is proposed that demonstrates iris classification which is made by combining the weighted average sum of the accuracy attained by various classifiers. This model is trained and tested on well- known iris datasets: Ubiris Version 2 (part1) and Ubiris Version 2 (part2), Casia Iris Interval. The paper resulted in the accuracy of the proposed system of Ensemble Learning on various epochs stating that the accuracy is directly dependent on the number of epochs as on Casia Iris Interval Dataset, the accuracy of the Ensemble model at epoch 10 (77.86%), epoch 30 (83.79%), epoch 50 (86.00%) and epoch 100 (87.24%) which is increasing as number of epochs increases. The paper also proves that the performance of the new system is better than the other base models. According to one of the dataset, Casia Iris Interval dataset, the proposed Ensemble Learning model’s accuracy on 100 epoch was 87.24%, which is significantly higher than the accuracy of the other base models, including DenseNet121 (70.88%), MobileNet (86.51%), InceptionV3 (63.61%), InceptionResNetV2 (34.09%), Xception (68.45%), and CNN (4.07%) respectively.