3D Assisted Deep Learning Face Recognition

Face recognition is a kind of biometric technology to identify or verify a person from the input facial information.In computer graphic and computer vision field,face recognition has been widely researched.Traditional face recognition algorithms use color images to identify persons.2D color face images are easy to obtain and suitable for deep learning method,which have promoted 2D face recognition a lot over the past few years.However there are some drawbacks of 2D face recognition,in which its accuracy highly depends on illumination,face pose and expression.In some specific scenarios,for instance,2D face recognition turns very unreliable when light dimmed.For one thing,the development of 3D face recognition technology can bring new turning points for face recognition.For another,with the rise of depth cameras,3D face information can aid face recognition task to improve its accuracy.Based on the above analysis,we propose a new data processing approach for color face images,where the 2D face image is aligned via the reconstructed 3D face model.First,given the input face image,we use an off-the-shelf parametric face reconstruction algorithm to reconstruct the 3D face mesh,which is the linear combination of facial identity and expression.Second,we only retain the identity by eliminating the expres-sion.Then the face mesh only with identity is parameterized onto the UV domain to get facial UV texture map.Finally,we train an elaborately-designed pseudo convolutional neural network with facial UV texture maps and face images as input.Extensive exper-iments and ablation studies are conducted to validate the effectiveness of our proposed algorithm and the superiority over traditional methods.Depth information and color information are combined together to get compound feature in general 3D face recognition algorithms.At present,many depth cameras in the market get the depth map based on the principle of structured light speckle imaging.Therefore we directly use the raw signal captured by structured light depth sensors which is the speckle image to preform face recognition.Inspired by some deep learning based stereo matching algorithm,we elaborately design a face recognition network whose input is the speckle image.Such an approach can fully exploit the rich information contained in speckle images.Extensive experiments and ablation studies are conducted to validate the effectiveness of our proposed framework and the superiority over face recognition on depth images.