A Deep Learning Based Iris Recognition System

Authentication by biometric verification is becoming increasingly common in security applications such as banking,border control,access control,and forensics.Apart from DNA identification,iris recognition is regarded as the most reliable and accurate biometric identification system,since irises are featured uniqueness,stability over time,large capacity,anti-counterfeiting,and non-contact acquisition.Compared with fingerprint recognition,iris recognition is more convenient and has better performance in anti-counterfeiting.Compared with face recognition,a hot topic in recent years,iris recognition is more reliable and stable.Therefore,since the 1990s,iris recognition systems emerged.Most Iris recognition systems follow Daugman's four-step model,namely,iris segmentation,normalization,feature extraction,and template matching.Early iris recognition systems mainly used iris features obtained by man-made extractors.These systems have bad performance on iris images with light reflection,blurring,or those acquired from distance.Improving the generalization ability of recognition system becomes the key to the development of iris recognition technology.With the successful application of deep learning in the field of computer vision,deep learning has been applied to iris recognition since 2015.Particularly,the UniNet architecture proposed by Z.Zhao et al.in 2017 achieved the best performance on the ND-IRIS-0405,CASIA.v4-distance,ITTD,and WVU Non-ideal datasets.This thesis develops an iris recognition system by applying deep learning.The main results are as follows.1.Iris segmentationInspired by the application of MTCNN to face detection,a cascaded convolutional network was designed for iris segmentation.Experiments show that our model performs excellently,even for iris images with occlusion,blur,light reflection,or eyeglass frame interference.Moreover,our model is able to detect many irises in one image.2.Iris feature extractionA semi-supervised iris feature extractor is designed based on a fully convolutional network,so as to generate spatially corresponding iris feature descriptors.The proposed framework outputs the feature map of the input and its mask simultaneously.By contrast,the UniNet should rely on the MaskNet,an additional network which is supervised trained,for the mask generating.The iris recognition system developed was trained on the CASIA-V4-Thousand dataset only.Without any further tuning,the system achieves high matching accuracy on other datasets.Thorough cross-datasets experiments on two databases suggest that the proposed framework outperforms the CrossDB model of UniNet.Experiments results indicate that our system is accurate and highly generalizable,and so is practical.