Research And Application Of The Key Technology Of Stereo Matching Based On Deep Learning

Nowadays,face recognition has been applied in every aspect of our life,such as face payment,attendance management and so on.Although two-dimensional face recognition has been very mature,it still has shortcomings of poor security and high misrecognition.Therefore,the focus is gradually shifting to 3D face recognition,how to acquire 3D face model is the basic knowledge.Three-dimensional acquisition based on binocular stereo vision has become a hot spot in this field due to its fast modeling speed and high accuracy,including two methods of active measurement and passive measurement.Passive 3D measurement technology is to capture images under natural light sources and restore 3D information of objects by matching the corresponding points in stereo images.How to improve the precision of stereo matching is the key.However,passive 3D measurement is difficult to meet the high-precision modeling scene,3D measurement based on fringe structure light encodes points on the surface of the object by projecting “phase features”,and then obtain high-precision 3D information through phase matching.Phase unwrapping is the key of this technology,which determining the precision of stereo matching based on the phase.Traditional stereo matching algorithms have relative accurate results on most objects,but it is easy to mismatch due to inability to extract effective features on textureless areas and reflective surfaces.Tradition phase unwrapping algorithms are difficult to extract accurate phase feature in the problem areas such as under-sampling and phase discontinuous region,so that the phase matching cannot be accurately achieved.In both active and passive stereo matching,it is easy to be mismatched in the problem area because the effective features cannot be extracted accurately.Convolutional neural network has powerful feature extraction ability for stereo matching task,so it becomes the mainstream of stereo matching research.The main research work and contributions of this paper are as follows:(1)A stereo matching algorithm based on dual attention mechanism is proposed to improve the mismatching problem of existing stereo matching algorithms on weak texture and reflective surface.In feature extraction network,dual attention module is added to extract features containing richer contextual information.The improved hourglass module is used in the cost aggregation network,which recovers the spatial position information in the deep layer of the network to accurately predict the disparity,can significantly overcome the mismatching problem of the reflection surface in the weak texture region.The MAE of Scene Flow dataset is0.87 px and the 3-px-error of KITTI is 1.80%.Furthermore,the precision of this algorithm on face1 dataset is better than algorithms,such as SGM and CFPNet,it can applied in 3D face construction under outdoor scenes.(2)A novel phase unwrapping algorithm based on multi-scale fusion is proposed,which can solve the difficulties of phase unwrapping in under-sampling and phase discontinuous region.In this algorithm,multi-scale features are fused by encoding and decoding structure,and semantic subnetworks are embedded in the decoder part to capture semantic information,so that phase unwrapping can be done quickly and accurately,especially in under-sampling and phase discontinuous regions.Compared with the time phase unwrapping algorithm that requires multiple fringe images(>=6),the algorithm in this paper only needs three fringe images to get the phase accuracy equivalent to the time phase unwrapping algorithm.Applying this algorithm to the face2 dataset,the RMSE is 0.0387,and the SSIM is as high as 0.9850,which can meet the needs of high-precision three-dimensional face construction;the RMSE on the mask dataset is 0.0273,and the SSIM is 0.9793.It can be proved that the algorithm has good generalization.(3)Two face datasets(face1 and face2)and one mask dataset(mask)are constructed.face1 contains 400 sets of samples,each set of samples includes binocular stereo image pairs and the disparity maps as ground truth values,which were used for the training and testing of stereo matching algorithms;face2 includes 50000 sets of samples,and each set of samples includes wrapped phase and the corresponding unwrapped phase as ground truth value,for training and testing of the phase unwrapping algorithms;the mask data set includes zodiac signs masks,twelve constellations masks and so on,with a total of 100 sets of samples,and each group of samples includes wrapped phase and the ground truth,which are used to test the phase unwrapping algorithm.This paper studies the key technologies of stereo matching.A stereo matching network is proposed to extract robust texture features for passive stereo matching to achieve 3D face construction under outdoor scenes;a phase unwrapping network is proposed to extract accurate phase features for active 3D modeling to achieve high-precision 3D face construction.