基于光照过滤的多模态活体检测  

Multimodal live detection based on light filtering

作　　者：高文杰 邵叱风 GAO Wenjie;SHAO Chifeng(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《哈尔滨商业大学学报(自然科学版)》2025年第2期176-183,共8页Journal of Harbin University of Commerce:Natural Sciences Edition

基　　金：国家自然科学基金(61572034);安徽高校与人工智能研究院协同创新项目(GXXT-2021-006).

摘　　要：针对人脸活体检测算法易受光照强度变化影响且在跨域检测时准确率较低等问题,提出了一种基于伪负样本生成并结合光照过滤模块的多模态活体检测模型LFFAS.同时采用RGB和LBP图像作为输入,挖掘活体人脸的特征信息.构造光照过滤模块,用以消除光照变化对检测结果的干扰.通过伪负样本生成的方法,构建活体人脸的隐空间,用以防御各种未知非活体攻击.实验结果表明,LFFAS模型在OCIM跨域测试上的半错误率(HTER)分别为14.0%、15.19%、16.38%、14.01%,相较于现有主流模型具有更好的检测性能.A multimodal face detection model,LFFAS,based on pseudo-negative sample generation combined with a light filtering module,was proposed to address the issues that face vivisection algorithms were susceptible to changes in light intensity and had low accuracy in cross-domain detection.RGB and LBP images were used as inputs to explore the feature information of vivisected faces.A light filtering module was constructed to eliminate the interference of lighting changes on the detection results.At the same time,the pseudo-negative sample generation method was employed to construct the hidden space of the living face,which helped defend against various unknown non-living body attacks.Experimental results showed that the half-total error rate(HTER)of the LFFAS model in the OCIM cross-domain test was 14.0%,15.19%,16.38%,and 14.01%,respectively,outperforming existing mainstream models.

关 键 词：活体检测 数据融合 LBP算子 光照过滤 伪负样本生成 域泛化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]