Research On Face Reconstruction Method Based On Image Transformation And Deep Learning Under Multi-pose Change

In recent years,positive face images have been widely used as an important identity feature,but in practical applications such as surveillance cameras,there may be a face that cannot accurately capture a subject,but it can be easier.Obtaining a number of non-positive faces of different poses of the same subject.Therefore,it is of great practical significance to accurately obtain the frontal face image of the subject through the multi-pose side face image of a certain subject.In view of the fact that most of the current situation can not reliably obtain the frontal face image of a subject,this thesis aims to reconstruct the frontal face image,and uses the multipose side face image as a sample to change from small posture to large posture change.The face of the face is reconstructed from the frontal face image,and the effect of different methods on reconstructing the frontal face image of the multi-pose side face image is explored.The main work of this thesis includes the following aspects:Firstly,Based on the face reconstruction of the face with small gestures,using a thin plate spline transformation and a moving least squares transformation method,selecting the pixels with the least distortion from a plurality of non-positive face images for synthesis,further utilizing the weighted average The method is to smooth the synthesized frontal face image so that the reconstructed frontal face image is as smooth as possible.Secondly,The problem of serious missing information on the side of the attitude deflection direction of the frontal face image reconstructed based on the image transformation method.A face reconstruction method based on the improved image fusion strategy under segmental affine transformation is proposed.In this method,the input non-positive face image is mapped to the frontal face template by segmental affine transformation to obtain the preliminary positive face;using the Poisson fusion idea,the fusion strategy is improved,and each transformation is performed according to a certain weight.The image blocks are selected from the subsequent frontal face images,and the image blocks are merged.This method effectively improves the problem of reconstructing missing face information.Thirdly,Based on the positive face reconstruction scene under the large attitude change,the positive face reconstruction method based on multi-task learning is proposed for the problem that the reconstructed positive face is not true enough and the detail information is not serious.On the basis of the stepwise reconstruction of the stack progressive auto-encoder to obtain the frontal face,considering the problem that the local feature information is easily lost during the face reconstruction process,the network optimization is performed,and the sparse automatic encoder based on partial feature expression is introduced.The stack progressive autoencoder is constructed with multi-task learning,and the local feature information of the input data is retained,thereby improving the reconstruction quality of the face image,so that the same person has the same degree of facial features in different postures.