| With the increasing demand for identity authentication,the iris recognition system is gradually applied to various scenarios.As a part of iris recognition,the accuracy of iris segmentation directly affects the accuracy of the recognition system,and its efficiency directly affects the performance of the recognition system.In this paper,the key problems of iris segmentation in different scenes are studied.Corresponding iris segmentation methods are proposed for different problems.The main research contents are as follows:(1)Iris segmentation for multiple types of acquisition devices: In many office scenes,there are many kinds of iris image acquisition devices,and iris images captured by different camera sensors differ in resolution,image quality,and gray level.However,most iris segmentation methods are typically designed for a single collection device,and their segmentation performance significantly decreases when processing the aforementioned types of iris images.To solve this problem,this paper proposes an iris segmentation network DMS-UNet(Dropblock and Modified Shortcut branch U-Net),which is suitable for multiple acquisition devices.In the encoder part,an improved shortcut branch structure is designed to reduce the loss of iris texture,edge,and fine-grained features.In the encoder and decoder part of the network,Dropblock technology is used to strengthen the learning of the lost feature information and improve the generalization ability of the network.(2)Iris segmentation for resource-constrained devices: Currently,high-precision deep neural networks usually rely on high computational complexity and large storage space,which cannot be directly deployed on some resource-constrained devices,such as edge mobile devices and mobile terminals.In order to balance the accuracy and efficiency of segmentation,this paper proposes an iris segmentation network ATT-LWNet(ATTention-Light Weight Net)suitable for resource-constrained devices.In the early stage of the network encoder,the spatial attention mechanism is used to capture multi-scale feature information without significantly increasing the computational cost,so as to improve the segmentation accuracy of the network.In the decoder part,the transposed convolution and interpolation methods are combined to effectively reduce the amount of calculation and storage space capacity of the network.(3)Iris segmentation for small sample datasets: In scenarios such as families,small laboratories,and small businesses,the limited number of collectors leads to fewer iris image samples.Iris segmentation networks with complex structures usually contain millions of parameters.Such networks tend to overfit on small sample data sets,thus reducing the accuracy of segmentation.In order to solve this problem,this paper proposes a Lightweight Iris Boundary Localization Network(LBL-Net)with a simple structure and very low parameters.In the encoder,the designed multi-scale context information extraction module can not only effectively fuse the information of different scales,but also reduce the interference of non-iris regions on the network.A multi-level feature information fusion module is designed between the encoder and decoder to fuse the deep and shallow features.This module can reduce the loss of feature information caused by the down-sampling phase of the network.DMS-UNet achieves good segmentation accuracy in cross-database test and hybrid database test,which proves that the network can accurately segment iris images collected by different devices.ATT-LWNet can greatly reduce the computational and hardware costs of the network under the condition of achieving high segmentation accuracy.Compared with the existing iris segmentation networks,LBL-Net achieves better segmentation accuracy with fewer parameters on small sample data sets. |
