Representation Learning For Low Illumination Face Detection And Recognition

Face recognition system has been widely used in the cooperative scenes,and can effectively carry out face recognition.However,there are still challenges in the face recognition system in uncooperative low illumination scenes.This dissertation focuses on two core tasks of face recognition system under low illumination:face detection and face recognition.This dissertation mainly uses the method of representation learning to eliminate the influence of low illumination by learning the representation with low illumination robustness.For face detection task,this dissertation uses illumination enhancement module to strengthen image representation and eliminate the influence brought by low illumination.For face recognition task,this dissertation uses disentangled representation learning method to learn face representation and illumination representation,so that face representation becomes illumination independent.Therefore,this dissertation proposes the representation learning for low illumination face detection and recognition method.The specific research content is as follows:1.This dissertation proposes an enhancement-based low illumination face detection method.In this method,illumination enhancement and face detection are integrated to enhance the representation in the process of representation extraction,so as to reduce the mismatch between image enhancement and face detection tasks.Specifically,the proposed method consists of two components,i.e.,representation enhancement and face detection.For representation enhancement,both global and local enhancement are explored.Global enhancement utilizes illumination constraints to enhance the overall image illumination,while local enhancement employs a generative adversarial network to improve illumination of face regions during training.Through combination of global and local enhancement,we reduce the gap between image enhancement and face detection,and optimize the two tasks jointly.Furthermore,regions of human body may help face detection,especially when face regions are small and blur.Therefore,we regard body region as context to assist face detection.Specifically,we use the context assistance module from face detection method Pyramidbox and optimize the body annonation with the aid of human structure prior knowledge.Experiments on the Dark Face dataset demonstrate the effectiveness of the proposed method.2.This dissertation proposes a low illumination face recognition method based on representation disentanglement.Inspired by Retinex theory,face images are disentangled into face-related and illumination-related information to learn face representation.Specifically,this dissertation uses a two-stream network to extract the face and illumination representation of the input image.At the image level,generative adversarial network is used to ensure the integrity and accuracy of face and illumination representation by reconstruction and cross reconstruction.At the representation level,through the analysis of influencing factors of illumination representation,it is found that the same illumination change has the same influence on illumination representation even for different faces.Therefore,a illumination offset loss is designed to constrain the illumination representation according to the invariance of the influence of illumination variation on the illumination representation.Finally,high-dimensional face representation is mapped to low-dimensional face representation by face classifier,which is used in face recognition task.In the experimental phase,the performance of the proposed method and related work is compared on four datasets:Yale Extend B,Peal R1,CASIA NIR-VIS 2.0,Oulu-CASIA NIR-VIS,which proves the effectiveness of the proposed method.In summary,this dissertation employs representation learning to mitigate the impact of low illumination on face detection and recognition tasks,thereby enhancing the performance of low illumination face detection and recognition.