Research On Deformation Measurement Methods With Feature Recognition

Deformation measurement is an essential approach to characterize and explore deformational behaviors of objects,with development of measurement techniques,modern optical measurement methods based on image and computer vision in photomechanics have been increasingly applied to fields such as mechanical,civil and aerospace engineering,etc.However,a mount of practical applications require that measurement systems work in uncontrollable environments,raising gigantic challenges to image and vision-based deformation measurement techniques.For that,this thesis aims to research on several key techniques in image and visual deformation measurement based on algorithms of image feature recognition:(1)A salient feature based inverse compositional Gauss-Newton(IC-GN)method with damping strategy is studied,which aims to compute trusted deformation data from natural texture feature or low-quality speckle feature.This method uses a feature detection technique to extract a full-field distributed salient features from the reference image as computation points,then a damped IC-GN algorithm is employed to track feature locations in a deformed image for evaluating displacement information at each point.Salient features and damping strategy enhance the stability of subset matching effectively,this makes it possible to estimate sub-pixel displacements reliably without having to depend on the initial value transferred from seed point(s).Finally,the measurement accuracy and stability of the proposed method is tested by an experiment.In addition,a domain-transformation based optimization method of displacement field is proposed,which not only helps to eliminate outliers and noises in a displacement field,but preserves the inherent displacement gradient attribution.(2)By using scale-invariant feature detection technique,an automatic,high-precision stereo calibration and correction method are proposed with using unconstrained scene data.For that,a new radial distortion model is introduced in the intrinsic parameter calibration,which directly maps image points from distorted image to undistorted image domain,so that the back-projection model is established with an analytical form;considering unconstrained scene features,the inverse depth parameterization is introduced in the back-projection model,leading the estimation of calibration parameters does not depend on strong geometric constraints such as planarity;finally,a bundle adjustment(BA)model for system calibration is proposed by measuring reprojected errors in the normalized image domain.Experiments show that the calibration model has performance comparable to that of commonly used methods.In order to address external parameter disturbance in measurement process,a concise realtime external parameter correction method is derived from the BA model,and several simulated and real experiments are conducted to explore the performance of this method.(3)An inverse compositional correlation bundle adjustment(ICC-BA)is proposed for threedimensional(3D)visual deformation measurement,aims to address the stability problem of stereo deformation measurement system in dynamic or uncontrolled environment.ICC-BA is established by combining the IC-GN subset matching and geometric BA algorithms,it is capable of jointly optimizing structure parameters of imaging system,feature depths and deformation parameters,allowing us to achieve optimal 3D reconstruction for cases with unstable system external parameters;by employing the Lie algebra representation for camera pose,the proposed ICC-BA could estimate camera poses in motion from an image sequence.In addition,a high concurrent depth estimation algorithm is proposed for the stereo measurement system with a fixed light path structure.Finally,experiments are conducted for validating the feasibility and effectiveness of each algorithm.(4)Based on deep learning techniques,and guided by supervised learning,a learnable deep neural network model for deformable image feature recognition is studied,which could achieve ultra-robust deformation measurement under bad conditions like big image noise,etc.A professional dataset construction method for model training is firstly proposed with using speckle pattern simulation and data augmentation,then a learnable deformed image feature recognition model is established based on convolutional neural networks and traditional subset matching algorithm.For validating the performance of this model,it is trained by a set of simulated images and tested by simulated speckle images and real deformed images.The simulation test shows that the model not only has the measurement accuracy consistent with the traditional method,but also predicted results are almost immune to noise;the actual test proves its generalization ability and the validity of measurement results in the actual measurement.