| Face recognition is a biometric technology based on facial feature information,which is one of the most attractive problems in the computer vision community.In practical application scenarios such as traffic monitoring driver's face recognition or electronic payment face recognition,face images are collected in outdoor environments,often with complex illumination variations.In order to solve the face recognition problems under complex illumination conditions,corresponding solutions are designed from two perspectives.On the one hand,an illumination recovery model is designed from the perspective of converting complex illumination variations into slight/moderate illumination variations to eliminate the influence of complex illumination variations on the face image.On the other hand,a specific loss function is designed to improve the discrimination of deep face features based on the effects of illumination variations.Firstly,in view of the problem that complex illumination variations will affect the face image,an illumination recovery model(GRIR)for face recognition with complex illumination variations is proposed.The GRIR model first uses the SVD algorithm to factorize the logarithm version of the face image with complex illumination variations.The singular values obtained can describe the illumination intensities for their corresponding SVD bases,which represents the facial features with different frequencies.Then,the singular values are normalized to generate a reference image and finally generated reference image based illumination recovery image is calculated by the gradient descent algorithm.The GRIR model can avoid the loss of inherent information on the face,and convert complex illumination variations into slight/moderate illumination variations,effectively improving the accuracy of face recognition with complex illumination variations.Secondly,in view of the problem of poor discrimination of traditional deep face features under complex illumination conditions,a face recognition method based on quadruplet-center loss(QCL)is proposed.The QCL method first combines the center loss with the quadruplet loss by replacing the negative sample with the center of deep features,and then uses a hard sample mining strategy.Only the negative sample center closest to the anchor samples are selected in each batch.Finally,based on the average distributions of the positive pair distance and the negative pair distance in each batch,a dynamic threshold is proposed to effectively prevent the model from the over-sampling or under-sampling problems.The QCL method comprehensively considers the relative distance and absolute distance between positive and negative sample pairs,thus improves the tightness and separation between faces.Therefore,the discrimination of deep features is strengthened.Experiments show that the prediction accuracy of face recognition with complex illumination variations can be greatly improved.Finally,in view of the problem of complex illumination variations during the training of the deep model,a deep face recognition method for complex illumination variations is proposed.The GRIR and QCL methods are combined into the GRIR-QCL method,where the GRIR model is used to unify the illumination of the training set and the test set,and then the QCL method is used to train the model.Then,an image enhancement method for simulating complex illumination variations(SCLV-IA)is proposed,thus generate a complex illumination variations(SCLV)dataset.Finally,this dataset is used to fine-tune the QCL pre-trained model.The GRIR-QCL method eliminates the impact of illumination variations on the face image from the perspective of preprocessing methods;the SCLV-IA method enhances the discrimination of deep face features to complex illumination variations from the perspective of transfer learning.Experiments show that both methods improve the face recognition rate with complex illumination variations.Compared with traditional illumination processing methods and deep learning methods,the proposed algorithms can obtain better recognition results on the Extended Yale B,CMU PIE,LFW,Mega Face,VGGFace2 test and YTF datasets. |
