基于注意力特征融合网络的手指静脉图像质量评价方法  

Finger vein image quality evaluation method based on attention feature fusion network

作　　者：迟云浩 杨璐 郭杰 郝凡昌 聂秀山 CHI Yunhao;YANG Lu;GUO Jie;HAO Fanchang;NIE Xiushan(School of Computer Science and Technology,Shandong Jianzhu University,Jinan 250101,Shandong,China)

机构地区：[1]山东建筑大学计算机科学与技术学院,山东济南250101

出　　处：《山东大学学报（工学版）》2023年第6期56-62,共7页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(62076151);山东省自然科学基金资助项目(ZR2021JQ26,ZR2022MF272,ZR2021QF119);山东省泰山学者资助项目(tsqn202211182)。

摘　　要：为充分挖掘质量特征以提高手指静脉图像质量评价的性能,提出一种基于注意力特征融合的深度可分离卷积网络,将其用于手指静脉图像质量评价。该方法主要包括静脉纹路提取、深度质量特征提取、注意力特征融合和图像质量类别预测等四个步骤。使用深度可分离卷积代替传统卷积,减少网络参数,使网络轻量化。使用注意力特征融合代替特征串联融合,从手指静脉灰度图像和手指静脉纹路图像中挖掘更具区分性的质量特征。考虑到目前没有公开手指静脉图像质量数据库,手工标注山东大学手指静脉公开库中图像的质量标记。试验结果表明,本研究提出的方法在手工标注数据库上的图像质量分类正确率为89.67%,图像质量评价性能优于现有手指静脉图像质量评价方法。In order to enhance image quality feature and further improve the accuracy of finger vein image quality evaluation,a deep separable convolution network based on attention feature fusion was proposed and used for finger vein image quality evaluation.Four steps including vein pattern extraction,deep quality feature extraction,attention feature fusion,and quality label prediction were involved in the proposed method.A deep separable convolution,not traditional convolution,was used for reducing network parameters and making the network lightweight.At the same time,attention feature fusion instead of feature concatenation was used to mine more discriminative quality features from grayscale and vein pattern images.Considering that there was no public finger vein database with quality labels,the quality labels of the images in the public finger vein database from Shandong University were manually annotated.The experimental results showed that the image quality classification rate of the proposed method on the annotated database was 89.76%,which outperformed the state-of-the-art finger vein image quality evaluation methods.

关 键 词：手指静脉识别 图像质量评价 卷积神经网络 注意力特征融合 深度可分离卷积 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]