基于图像深度先验和鲁棒马尔可夫随机场的有遮挡人脸识别  

Occluded Face Recognition Based on Deep Image Prior and Robust Markov Random Field

作　　者：李小薪 丁伟杰 方怡[1] 张远成 王琦晖 LI Xiaoxin;DING Weijie;FANG Yi;ZHANG Yuancheng;WANG Qihui(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;Department of Computer and Information Security,Zhejiang Police College,Hangzhou 310053,China;Qianjiang College of Hangzhou Normal University,Hangzhou 311121,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]浙江警察学院计算机与信息安全系,杭州310053 [3]杭州师范大学钱江学院,杭州311121

出　　处：《计算机科学》2024年第7期244-256,共13页Computer Science

基　　金：浙江省自然科学基金(LGF22F020027);公安部科技强警基础工作专项项目(2020GABJC35);国家自然科学基金(62271448)。

摘　　要：由遮挡所引发的测试数据和训练数据之间的差异,是人脸识别技术面临的重要挑战。现有的基于深度神经网络的有遮挡人脸识别方法大多需要使用大规模的有遮挡的人脸图像来训练网络模型。然而,现实世界中的任何外界物体都有可能成为遮挡,有限的训练集数据很难穷尽所有的可能性,并且使用大规模的有遮挡人脸图像训练网络模型的做法与人类视觉机制是相违背的,人眼对于遮挡区域的感知在本质上与遮挡本身并没有关系,仅依赖于无遮挡的人脸图像。为了模拟人类视觉的遮挡检测机制,将图像深度先验和鲁棒马尔可夫随机场模型结合起来,构建基于小样本数据的遮挡检测模型DIP-rMRF,并提出了一致性零填充方法以有效利用DIP-rMRF的遮挡检测结果进行后续的人脸识别。在Extended Yale B,AR和LFW这3个人脸数据库上,针对VGGFace, LCNN,PCANet, SphereFace, InterpretFR,FROM这6种CNN模型的实验结果表明,DIP-rMRF能够有效地处理遮挡以及由极端光照所引发的“类遮挡”,从而极大提升现有的深度神经网络模型对有遮挡人脸识别的性能。The occlusion-caused difference between test and training images is one of the most challenging issues for real-world face recognition system.Most of the existing occluded face recognition methods based on deep neural networks(DNNs)need to use large-scale occluded face images to train network models.However,any external object in the real world might become occlusions,and limited training data cannot exhaust all possible objects.Also,using large-scale occluded face images to train networks violates the human visual mechanism,the human eyes detect occlusions by only using small-scale unoccluded face images without seeing any occlusions.In order to simulate the occlusion detection mechanism of human vision,we combine the deep image prior with the robust Markov random field model to construct a novel occlusion detection model,namely DIP-rMRF,based on small-scale data,and propose a uniform zero filling method to effectively utilize the occlusion detection results of DIP-rMRF.Experimental results of six advanced DNN-based face recognitions methods,including VGGFace,LCNN,PCANet,SphereFace,InterpretFR and FROM,on three face datasets,including Extended Yale B,AR and LFW,show that DIP-rMRF can effectively preprocess face images with occlusions and quasi-occlusions caused by extreme illuminations,and greatly improve the performance of the existing DNN models for face recognition with occlusion.

关 键 词：有遮挡人脸识别 图像深度先验 鲁棒马尔可夫随机场 一致性零填充 结构误差度量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]