Low-Resolution And Occluded Face Recognition Enhanced By Privileged Information

Face recognition is one of the important research topics in the field of computer vision.The existing general face recognition method is aimed at high-quality face images.But the realistic application scenarios are mostly uncoordinated and uncontrolled,and the captured face images are low-quality.At this time,the general face recognition methods are difficult to achieve robust generalization performance,which limits the accuracy of face recognition results.The thesis focuses on two issues in uncoordinated and uncontrolled environments:low-resolution face recognition and occluded face recognition.Considering that the introduction of privileged information can assist model training,this thesis adopts a training framework based on privileged information.For low-resolution face recognition,this thesis proposes to use high-resolution face images as privileged information to guide the learning of low-resolution face recognition model;For occluded face recognition,this thesis proposes to use non-occluded face images as privileged information to assist the learning of occluded face recognition model.Therefore,this thesis proposes a low-resolution face recognition and occluded face recognition method enhanced by privileged information.The details are as follows:1.This thesis proposes a single-stage dual-spatial supervision method for lowresolution face recognition.This method utilizes high-resolution face images as privileged information to guide the learning of low-resolution face recognition model in feature map space and representation space.In the feature map space,this method introduces the similarity constraint and reconstruction loss of multiple intermediate layer feature maps.In the representation space,this method proposes a supervised auxiliary contrastive loss function to strengthen the approximation between high-resolution and low-resolution representation distributions.In order to adapt to the input of multiple resolutions,this method proposes a oneto-many matching strategy and an adaptive weight adjustment strategy.Experimental results on three synthetic low-resolution face databases and three realistic low-resolution face databases demonstrate the effectiveness of this method.2.This thesis proposes a two-stage multi-scale resolution-adaptive low-resolution face recognition method.This method improves the first method through twostage training manner.Stage one is the multi-scale distillation stage,which aligns the multi-scale feature map distribution and representation distribution between high-resolution and low-resolution through pixel value error,affinity matrix distillation and mutual information maximization.This method proposes a simpleto-complex curriculum learning strategy.Stage two is the multi-resolution clustering stage.Stage two proposes a multi-resolution contrastive loss function.The fine-tuning is used to make the final model maintain good generalization performance in real low-resolution scenes.Experimental results on three synthetic low-resolution face databases and three realistic low-resolution face databases demonstrate the effectiveness of this method.3.This thesis constructs a large-scale occluded face database and proposes an occluded face recognition method based on representation alignment and semantic segmentation.This database has a data volume of one million.Each occluded face image contains a corresponding occlusion mask.This method uses nonoccluded face images as privileged information to assist the training of the occluded face recognition model.This method introduces the similarity constraint and KL divergence to align occluded face representations and non-occluded face representations.Meanwhile,this method uses the occlusion mask label to guide the learning of the mask matrix.Through the weighting of the mask matrix,the weight of the non-occlusion areas are enhanced and the interference of occlusion objects are reduced.Experimental results on three synthetic occluded face databases and a realistic occluded face database demonstrate the effectiveness of this method.In summary,this thesis adopts the framework based on privileged information learning,and uses high-resolution face images and non-occluded face images to assist the representation learning of low-resolution face images and occluded face images.This thesis provides a new research idea for face recognition in the uncoordinated and uncontrolled environment,and improves the effect of low-resolution face recognition and occluded face recognition.