摘要
针对低质量虹膜图像分割精度低问题,提出一种基于SRN-UNet(SEResNext-UNet)的虹膜分割算法。在编码阶段增加SE-ResNext模块,该模块在RexNext模块后级联SENet(Squeeze-andExcitation Network)模块,以聚焦目标特征,在不增加网络参数的情况下提升分割精度和网络性能。在解码阶段的上采样层降低模型参数量,以提升训练速度。为解决图像类别不均衡问题,结合两种损失函数Dice Loss和Focal Loss训练SRN-UNet网络。基于CASIA-Iris数据集和自建的低质量虹膜图像数据集的实验结果表明,与其它算法相比,所提算法在视觉效果和客观评价指标上均有较好分割效果。与U-Net算法相比,所提算法的平均交并比、F1分数与精确率分别提升了4.20%、2.27%、5.38%,且运行速度高于U-Net。
In recent years,iris recognition has been widely used in various fields.Iris segmentation is the most critical step in the iris recognition process.The accuracy of the iris segmentation algorithm directly affects the performance of the entire iris recognition system.In this study,an iris image segmentation algorithm SRN-UNet(SeResNext-UNet)is proposed to solve the problem of low segmentation accuracy for segmenting low-quality iris images.In the coding stage,the SE-ResNext module is added,which is cascaded with the SENet(Squeeze-and-Excitation Network)module after the RexNext module.The ResNext module can improve the network performance without increasing the network parameters;the SENet module builds a network model from the perspective of feature channel correlation through squeeze,excitation,and weight redistribution.For low-quality iris images,the SENet uses global information to selectively emphasize informative features and suppress less useful ones,and improve the accuracy of iris segmentation.In the up-sampling layer of the decoding stage,the amount of model parameters is reduced to increase the training speed.In order to solve the problem of image category imbalance,the SRN-UNet is trained by combining the Focal loss function and the Dice loss function.Among them,the Focal loss function can reduce the weight of easy-to-classify samples,make the model pay more attention to the training of difficult samples,and guide the network to retain complex boundary details;the Dice loss function can solve the problem of pixel category imbalance and alleviate the noise caused by the Focal loss function.Experimental results based on CASIA-Iris dataset and self-built low-quality iris image dataset show that compared with other algorithms,the proposed algorithm has better segmentation effects in terms of visual effects and objective evaluation indicators.Among them,the Mean Intersection Over Union of the proposed algorithm reached 95.19%,the F1 score reached 97.48%,and the Precision reached 97.82%.Compared with U-Net,the Mean Intersection Over Union,F1 score and Precision of proposed algorithm have increased by 4.20%,2.27%,and 5.38%respectively,and the algorithm is faster than U-Net.
作者
田会娟
翟佳豪
柳建新
刘嘉伟
邓琳琳
TIAN Huijuan;ZHAI Jiahao;LIU Jianxin;LIU Jiawei;DENG Linlin(Tianjin Key Laboratory of Optoelectronic Detection Technology and System,School of Electrical and Electronics Engineering,Tiangong University,Tianjin 300387,China;Engineering Research Center of Ministry of Education on High Power Solid State Lighting Application System,Tianjin 300387,China;Tianjin Chengke Transmission Electromechanical Technology Co.,Ltd.,Tianjin 300384,China)
出处
《光子学报》
EI
CAS
CSCD
北大核心
2022年第2期241-249,共9页
Acta Photonica Sinica
基金
国家自然科学基金(No.61504095)
天津市科技计划(No.18ZXCLGX00090)。