结合反射图像的双流多层次融合人脸活体检测  

Face Anti-Spoofing Using Dual-Stream Multi-Level Fusion and Reflection Images

作　　者：王鑫 黄睿[1] WANG Xin;HUANG Rui(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《计算机工程与应用》2023年第13期238-246,共9页Computer Engineering and Applications

摘　　要：针对从RGB图像提取到的特征对光照敏感,导致人脸活体检测模型泛化性能较差的问题,提出一种结合反射图像的双流多层次融合检测(face anti-spoofing using dual-stream multi-level fusion and reflection images,DMF-RI)算法。对RGB图像进行带色彩恢复的多尺度Retinex增强,获得反射图,并分别提取原图和反射图的低、中、高多层次深度特征;通过双流多层次特征融合模块(dual multi-level feature fusion,DMFF)实现不同层次不同特征的有效融合;联合基于二值掩码的像素级监督和基于二值标签的二分类监督训练网络模型。在CASIA-FASD、Replay-Attack、MSU-MFSD、OULU-NPU和SiW这5个人脸活体检测数据集上的多组实验表明,所提算法通过多层次融合RGB图像和MSRCR图像的深度特征,能提取人脸中更为本质的特征信息,在复杂背景条件下表现出较强的鲁棒性和泛化能力。The features extracted from RGB images are sensitive to illumination and lead to poor model generalization performance of face anti-spoofing algorithms.To deal with the problem,a face anti-spoofing method using dual-stream multi-level fusion and reflection images is proposed.Firstly,the RGB images are enhanced through multi-scale retinex with color restoration(MSRCR),and the low,medium and high-level deep features of the RGB and resulting reflected images are extracted.Then,an effective fusion of different features at different levels is achieved through the dual multilevel feature fusion(DMFF)module.Finally,the network model is trained by joint pixel-wise binary supervision and binary supervision based binary label.Extensive experiments on five benchmark datasets,including CASIA-FASD,Replay-Attack,MSU-MFSD,OULU-NPU and SiW,show the better robustness and generalization capability of the proposed method which can fuse the multi-level features of RGB and MSRCR images and extract more intrinsic features of faces in the complex environments.

关 键 词：人脸活体检测 光照鲁棒性 多层次特征 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]