A Face Frontalization Method Based On 3D Reconstruction

The face recognition technology is regarded as one of the key technologies in the field of AI(artificial intelligence).Due to its advantages different from the traditional biological recognition technology,it plays an important role in the security protection area,and it has been rapidly expanded from single scenario to various types of hardware and different industries.At present,the application scenarios of face recognition mainly involve image collection under the cooperation of user,but in reality,most application scenarios cannot obtain cooperation of the collected object.The collection of face image information is often associated with uncertainty and randomness,and the face information captured by camera will present distortion,inclination and other interferences.Under this state,if the face image is directly input into the face recognition system or related application,the recognition and comparison cannot be conducted smoothly,in terms of recognition speed and rate.In order to solve this problem,we apply facial frontalization to facial recognition system.In order to address this problem,this paper proposes a face frontalization algorithm based on 3D face reconstruction for frontalization of input non-frontal face images.The deep learning approach is employed to train the network model to complete face detection,facial key point detection on face and other steps,the graphics method is used for affine transformation,and the frontalization of non-frontal face images is basically realized under normal conditions.This paper mainly consists of the following works:1.The status quo of current researches on face recognition and frontalization and the challenges faced in the application scenarios are summarized.The face frontalization procedure and related steps are reviewed,the main research results and methods for face detection,facial key point detection,face frontalization and face reconstruction are introduced,and the mainstream algorithms and network models are analyzed and compared.2.On the basis of detailed analysis of current face frontalization algorithms,this paper proposes a face frontalization method based on 3D reconstruction.First of all,this algorithm conducts preprocessing of input face image,the deep learning method is used to train the network model,and different network models are adopted in different stages to obtain facial key point information of face.In the meantime,by combining the network model with affine transformation,the obtained 2D facial key point information of face is used to calculate the rotation matrix between the 2D face image and the actual face model,and generate the frontal face model corresponding to the input face image which can assist further recognition.The experimental results show that the face image generated with this method presents identity consistency with that of the facial image,which can help improve the accuracy of side face recognition by the face recognition system.