采用双流网络结构的虹膜卷缩轮检测方法
Iris curl wheel detection method based on two-stream network structure
张波 1王昌鹏1
作者信息
- 1. 沈阳化工大学计算机科学与技术学院,辽宁沈阳 110142;沈阳化工大学辽宁省化工过程工业智能化技术重点实验室,辽宁沈阳 110142
- 折叠
摘要
针对虹膜卷缩轮检测易受干扰而导致边界定位波动的问题,提出一种基于双流网络结构的虹膜卷缩轮检测方法.在归一化图像中,利用双流网络结构分别提取结构特征和纹理特征.融合模块融合来自不同流的特征,得到初定位区域.对该区域进行滤波操作,去除噪声干扰.利用边缘梯度算子检测、提取卷缩轮.实验结果表明,该方法正确检出率为91.2％,边缘定位AP值为0.706,平均检测速度为2.3秒/幅,相比其它算法,存在复杂干扰的情况下,保证检出率和检测速度的基础上,降低了边界波动.
Abstract
To address the problem that the iris curl wheel position detection is susceptible to fluctuations in curl wheel boundary localization due to interference.An iris curl wheel detection method was proposed based on a two-stream network structure.Structural and texture features were extracted separately in the normalized iris image.A custom fusion module adaptively fused the features from different streams to obtain an initial localization region of the iris curl wheel boundary.The region was subjec-ted to a filtering operation to remove interference.The edge gradient operator detected the initial localization region to extract the iris curl wheel boundary.Experimental results show that the method has a correct detection rate of 91.2％,an edge localization AP of 0.706,and an average detection speed of 2.3 seconds/frame.Compared with other algorithms,the detection rate and speed are guaranteed based on further reduction of boundary fluctuations in the presence of complex interference factors.
关键词
虹膜纹理/图像处理/虹膜卷缩轮/归一化图像/残差分割网络/纹理检测/边缘梯度算子Key words
iris texture/image processing/iris curl wheel/normalized image/ResSegnet model/texture detection/edge gradient operator引用本文复制引用
基金项目
辽宁省博士科研启动基金(2019-BS-191)
辽宁省教育厅科研项目(LJ2020023)
出版年
2024
NETL
NSTL
维普
万方数据