Hierarchical knowledge distillation decoupling network for low-resolution face recognition algorithm

Zhong Rui†， Song Yafeng， Zhou Xiaokang （SchoolofMathematicsamp;Computer Science，GannanNormal University，Ganzhou Jiangxi341ooo，China）

Abstract：Alargenumberoffacial detail features arelostinlow-resolutionface images，which makes therecognitionrateof manyclassical facerecognition models with goodperformancedecrease sharplyToaddressthisproblem，thispaperproposed ahierarchicalknowledgedistilationdecoupling（HKDD）network.Firstly，theconvolutionallayersoftheteachernetworkand thestudentnetworkperformed hierarchicalfeaturedistilltiontoenhancethestudentnetwork’sfeaturedescriptionabilityfor low-resolutionsamples.Thisdistilltionensuredthatthelow-resolutionfacefeatures extractedbytheintermediatelayersof the studentnetworkcloselyapproximatedthehigh-resolutionfacefeatures extractedbytheintermediatelayersof theteachernetwork，effectivelytransferingthepowerfulfeaturedescriptionabilityoftheteahernetwork’sintermediatelayerstothestudent network.Subsequently，the softmax layers of theteachernetwork andthe student network performed decoupling distlation， spliting thedistillaionlossattesofmaxlayerintotargetcassdistillationlossndon-targetclassdistillationlosedecoupling distilationcould playaguidingroleinthetrainingof student networksbyleveragingthesuppessdnon-targetdistillationloss，sothatthesudentnetworkcouldlearntheclasificationabilityofuniversalfacialfeaturesundertheguidanceof theteachernetwork，therebyensuringthatthestudentnetworkhasstrongclasificationabilityinunrestrictedscenes.Finally， byconducting validation experiments on several low-resolution face datasets，such as TinyFace and QMUL-SurvFace，the HKDD model demonstrates superior performance intermsof recognitionrateandreal-time performancecompared tootherrepresentativelow-resolutionfacerecognition models.Theexperimentalresultsconfirmtheefectivenessof HKDD in handling low-resolution face recognition tasks.

Keywords：low-resolutionfacerecognition；hierarchicalknowledgedistilationdecouplingnetwork；hierarchicalfeaturedis tillation；decoupling distillation；unrestricted scenes

0 引言

目前人脸识别技术已广泛应用于门禁、考勤、支付以及安防监控等场景。然而在以上应用场景中很容易因为距离远而采集到低分辨率人脸图像，这类人脸图像中丢失了大量面部细节特征，同时各种干扰因素（如：姿态偏转、表情变化、光照变化和遮挡等）的叠加，使得所采集的人脸图像与训练集中的高分辨率人脸图像间存在很大的分布差异。在这样复杂的应用场景中，即使是采用目前先进的人脸识别算法也很难取得令人满意的识别精度，因此在复杂的非限制性场景中对低分辨率人脸进行准确识别是一项极具挑战性的任务。目前主要的解决方案可以分为统一特征空间映射法[1]、超分辨率重建法[2.3]和知识蒸馏法[4.5]三类，其中超分辨率重建法是将超分辨率重建与人脸识别两个过程进行集成，能够在一定程度上提升低分辨率人脸识别算法的识别率。但是图像的超分辨率重建仅仅是在视觉感官层面对像素进行了复原，而面部识别更重要的是在特征层面上对特征进行复原，该类型算法很难显著提升算法的识别率，且还需要耗费大量的计算资源，将导致算法的实时性不足[。统一特征投影法是将不同分辨率的人脸特征投影至同一个特征空间进行面部识别，采用该方式忽略了领域内部的结构差异性，也就是说，这种全局特征投影变换方法只关注于整体的特征差异，没有关注局部的特征差异，该类型方法只能够在一定程度上提高低分辨率人脸的识别率。……