Research On Three-dimensional Face Reconstruction And Recognition Method Based On Structured Light Measurement

With the development of information technology,face detection and identification technology based on biological individual characteristics has attracted more and more attention and research from researchers.Relying on the progress of deep learning technology,two-dimensional face recognition technology based on grayscale or color image information has significantly promoted the development of society as a whole.However,because the measured face is often unconstrained in the real recognition scene,the captured two-dimensional face information is easily affected by factors such as ambient light source,shooting posture,and makeup,thereby reducing its recognition performance.With the rapid development of 3D imaging equipment,the obtained 3D face data contains robust facial geometric information and depth information,which can overcome the limitations of 2D face recognition.For extreme non-controllable recognition scenarios,the performance of the entire3 D face recognition system will still be affected by complex interference factors,so that the optimal recognition rate cannot be achieved.Obtaining high-quality 3D face data is the essential foundation of the entire 3D recognition process,and excellent 3D face recognition algorithms can extract discriminative facial features from limited or disturbed 3D face data and accurately recognize individuals.Based on this,this dissertation mainly studies the following.First,a three-dimensional face recognition system based on encoded structured light is designed and implemented.The three-dimensional face data obtained by a threedimensional scanning device is more representative of the true essence of the face structure,which can effectively overcome the limitations of two-dimensional face data.The hardware of the 3D face recognition system studied in this dissertation consists of a camera,a DLP projector,and a computer.After projecting and collecting the sequence of fringe patterns,the system obtains a 3D point cloud of the face through algorithmic steps such as wrapped phase calculation,phase unwrapping,and phase-3D conversion.Finally,the system performs face recognition using preprocessing techniques and deep feature extraction.Secondly,this dissertation proposes a mechanism for removing invalid points from the face fringe patterns based on the error energy function.This mechanism is developed by utilizing the characteristic that the effective points of the captured cofrequency fringe image sequence are approximately distributed on the ideal cosine curve.This method involves constructing a relationship between the captured pixel values of the co-frequency pattern and the ideal cosine curve as the error energy.The proposed mechanism eliminates invalid points with high error energy by using an adaptive error threshold selection method.Compared to the popular fringe invalid point removal algorithm,this method can increase the mean intersection over union(MIo U)value by up to 6.25% and reduce the misclassification error(ME)value by at least 25%.And the method is not sensitive to object reflectivity or intensity modulation,and can be applied to different measurement environments.Thirdly,this dissertation proposes an unwrapped phase error correction algorithm based on multiple linear regression analysis,which is developed by utilizing the low rank characteristics of the overall information of the unwrapped phase map.This dissertation uses multiple linear regression analysis(MLRA)to obtain the regression plane of the unwrapped phase map based on its low ranking characteristics.The proposed algorithm sets the tolerance marker coarse error,marks the random error using the median filtering method,and finally corrects all the marked phase error points.The proposed method is based on phase unwrapping and can be applied without being limited by the hardware system or the principle of the phase unwrapping algorithm.Compared to popular phase correction algorithms,this method can effectively and robustly correct phase unwrapping errors in dense transitions.Fourth,according to the problem that the performance of 3D face recognition will be affected by face occlusion,this dissertation proposes a partially occlusion 3D face recognition network based on multi-feature combination threshold technology.This method characterizes the three-dimensional face as a multi-feature attribute map,compares it with the corresponding mixed average face to obtain a face mask template that focuses on different occlusion areas,and then uses this template to calculate the respective masks in order to locate the face occlusion area.After removing the occlusion,it is input into the convolutional neural network for recognition.The average Top-1 recognition rate of this method for Bosphorus database was 99.52%.The average Top-1 recognition rate for UMB-DB databases was 95.08%.Compared with the popular identification methods,it increased by 0.61% and 5.64%,respectively.Fifth,aiming at the problem that the number of 3D face data is small regardless of the type or its own category,an active face mask coding scheme is proposed to generate a large-scale labeled 3D face training set.This method divides the existing expression face into a 4×4 grid area size and generates face training samples with different missing face data by removing the face area block by block through different mask matrices.This method has a 100% recognition rate of Top-1 for Bosphorus database and UMBDB database for unobstructed posture face and expression face.For the Bosphorus database,the Top-1 recognition rate of the shielded face after removing the occlusion increased from 97.07% to 98.94%.In this dissertation,a three-dimensional imaging device based on structured light is studied and built for reconstructing three-dimensional faces.The proposed research addresses the shortcomings of existing three-dimensional face recognition systems by improving the quality of reconstructed three-dimensional faces,reducing phase blur during face unwrapping,increasing the accuracy of face recognition,and improving the diversity of three-dimensional face data samples.The proposed algorithms are extensively verified through experiments.Experimental results show that the algorithm proposed in this dissertation can effectively improve the performance of key technologies in the 3D face recognition system.