Research Of Iris Segmentation Based On Deep Learning

Iris segmentation is a key step of iris recognition.With the higher requirement of iris recognition applications,iris segmentation algorithms are facing more and more challenges in complex unrestricted scenes and cross-device applications.In order to improve the performance of iris segmentation,two iris segmentation algorithms are proposed,which are based on multi-task convolution neural network and fully convolution neural network.The iris segmentation algorithm based on multi-task convolution network determines the iris region by detecting the key points of iris boundary.Firstly,the radial symmetry transformation algorithm is used to determine the pupil position and centralize the image with the pupil position.Secondly,a lightweight key point detection convolution network is proposed.Four key point detection related tasks are designed:light occlusion,frame occlusion,severe eyelid occlusion,and eyes opened completely.The multi-task learning strategy is utilized for assisting the key point detection task to improve the generalization ability of the model.Finally,the boundary curve is estimated according to the key points and the closed area surrounded by the boundary curve is taken as the iris region.In C ASIA-Iris-Thousand database,the average location error of boundary point detection is 5.35%.The pixel-level accuracy and average intersection over union ratio of iris segmentation are 99.96%and 97.92%respectively and the model size is 7.32M.The key point detection network of an image takes 23.0ms.The iris segmentation method based on fully convolution network determines the iris region by pixel-by-pixel classification.Firstly,a lightweight fully convolution iris segmentation network is designed based on prior knowledge.Secondly,without increasing the network parameters,four optimization algorithms for iris segmentation are proposed:(1)Aiming at the imbalance of iris image pixel classification,the weighted loss makes the iris region pixel more concerned in the process of network training,which improves the iris region segmentation accuracy significantly.(2)In order to combine the deep and shallow feature information effectively,a multi-level dilated convolution dense connection structure is designed,which makes the network enhance the effect of detail segmentation.(3)By employing the parallel training optimization strategy of multi-scale image pyramid and multi-supervised signal,the adaptability of the network to the scale change of iris target is enhanced and a robust segmentation model for the scale change is obtained.(4)Based on the idea of generative adversarial networks,the prediction output of the iris image segmentation network and the real label are put into the discriminator.In the process of zero-sum game between segmentation network and discriminator,the enhanced segmentation network produces more fine segmentation results for iris image.Experiments show that,combined with four optimization items,the pixel-level accuracy and average intersection over union ratio of the algorithm are 99.30%and 95.35%respectively in UBIRIS.v2 database,99.66%and 96.75%respectively in CASIA-IrisThousand database and the model size is only 6.2M.An image is segmented using 41.56ms and 63.03ms in UBIRIS.v2 and CASIA-Iris-Thousand database respectively.Finally,this paper extends the above-mentioned GAN iris segmentation to unsupervised iris image segmentation across devices.The visible iris image with label is regarded as the source domain and the knowledge learned in the source domain is transferred to the near infrared iris library without label in the target domain.For the purpose of illustrating the effectiveness of the algorithm,UBIRIS.v2 database is used as the source domain and CASIA-Iris-Thousand database is used as the target domain.The segmentation results are cross-device unsupervised.The pixel-level accuracy is 98.24%in target domain and 99.24%in source domain.This shows that the algorithm not only can segment the unlabeled iris image,but also has good generalization in cross-device images.