Research On Accurate 3D Reconstruction Method Based On Fringe Projection

The 3D reconstruction method based on fringe projection has the advantages of non-contact and high flexibility.And it has become one of the most commonly used 3D reconstruction methods for its higher accuracy compared with other 3D reconstruction methods.However,in the process of practical application,the results will have a large number of noise points due to many factors.At the same time,multiple additional fringe patterns with different frequencies are needed for obtaining accurate reconstruction results,which limits the applicability of the method.Based on this,the main work of this paper is as follows:(1)We propose a phase correction method based on the characteristics of phase distribution.There are noise points in the phase distribution due to the difficulty of displaying the fringe patterns in the background,shadow of object and other areas.And there are jump errors in the phase distribution for the rounding operation used in the phase unwrapping process.In order to remove the noise points and the jump errors,we propose a phase correction method based on the absolute phase of each row is monotonically increasing and consists of several continuous and monotonically increasing curves.Firstly,the absolute phase of each row is divided into several phase curves according to the phase continuity.Then each curve is classified as normal,noisy or jumpy curve depending on its position relationship with its neighboring curves.At last,each curve is processed based on the classification result to obtain an absolute phase without noise points and jump errors.(2)We propose a phase stereo match method combined with curve fitting.To reduce the influence of the smoothed regions in the absolute phase on the reconstruction results,a phase stereo match method combined with curve fitting is proposed.The method first identifies the phase smoothing regions based on the number of candidate matching points,and then applies curve fitting in the smoothing regions to mitigate the influence of such regions on the reconstruction accuracy.In addition,when the phase-smoothed region overlaps with the occluded region,a matching error is introduced.To address this issue,this paper constructs distance parameters in terms of coordinates of pixel and its disparity,and then refines the disparity map based on the average distance between a pixel point and each pixel point in its neighborhood to further improve the reconstruction accuracy.(3)We propose a single-frequency and accurate phase unwrapping method using deep learning.The conventional methods are difficult to achieve high-precision and singlefrequency phase unwrapping,i.e.,it is difficult to obtain high-quality phase distribution without the help of fringe patterns at extra frequencies.To solve this problem,we propose a single-frequency and accurate phase unwrapping method using deep learning.This method treats the phase unwrapping as a semantic segmentation problem based on the quantitative relationship between the wrapped phase and the absolute phase.Furthermore,we propose a multi-scale fusion segmentation network model named as VRNet,which takes the wrapped phase map as the input,extracts the semantic information in the wrapped phase map through several convolutional and pooling layers,and generates feature maps at multiple scales simultaneously.The feature map is then processed by the stacked feature fusion module to recover the spatial information of each scale in turn.At last,the fringe order of each pixel is outputted.Extensive experiments show that the proposed method can recover the 3D shape of the object with high accuracy.And the standard deviation between the reconstruction result of the complex part FSW inverter and the standard model can be less than 0.1 mm.In addition,the single-frequency phase unwrapping method using deep learning can reduce the number of fringe patterns from 12 to 4,and can obtain high-quality absolute phase at the same time.The average prediction accuracy is as high as 99.13%in scenes composed of multiple sculptures,which is better than other single-frequency phase unwrapping methods.Based on this,we design and implement a 3D reconstruction system based on fringe projection,which supports operations including image loading,deformed fringe pattern analysis,phase unwrapping by multi-frequency heterodyne method,single-frequency phase unwrapping,point cloud generation,etc.