基于gma光流网络的二维变形测量数字图像相关法初值估计

Abstract

Objective In the twodimensional deformation measurement of speckle images, the initial value estimation of the digital image correlation method exerts great influence on the computational efficiency and accuracy of algorithms. The calculation accuracy and speed of subpixel displacement iterative search algorithms in digital image correlation methods depend on whether the initial value estimation provided by the integral pixel displacement calculation is reasonable or not, and its convergence radius is generally in the range of several pixels. Therefore, the initial value estimation provided in the integer pixel displacement search phase should be as close to the real value as possible to ensure that the iterative algorithm can converge quickly and accurately, otherwise, it may converge slowly or even fail in the iterative process. The traditional initial value estimation methods including the humancomputer interaction method, Fourier transform method, and feature matching method, have some problems such as slow calculation speed and low calculation accuracy in the face of large deformation measurement and unclear speckle image features. Recently, the optical flow estimation network models in deep learning feature fast calculation speed, high calculation accuracy, and strong generalization in predicting motion displacement. We introduce the optical flow estimation network model in deep learning into the digital image correlation method and employ the displacement field predicted by the optical flow network as the initial value of the sub -1310002 -12 研究论文 第 43 卷 第 13 期/2023 年 7 月/光学学报 pixel iterative algorithm. Finally, the inverse compositional Gauss -Newton method is adopted to calculate the displacement field of speckle deformation images. We hope that the strategy of introducing deep learning into the digital image correlation method can provide a new idea for speckle deformation measurement.Methods First of all, we compare the calculation accuracy of several optical flow network models of FlowNet2, PWC -Net, RAFT, GMA, SeparableFlow, GMFlow, and FlowFormer, which have excellent performance on MPI Sintel test datasets on speckle images. Considering the calculation time, model size, and calculation accuracy, the GMA network model is chosen to provide initial value estimation for subpixel iterative algorithms. Then, a feature sampling module is added to the model for solving the problem that the GMA network needs to occupy a lot of GPU resources in highresolution speckle images, which can effectively reduce the occupation of GPU memory by adjusting the sampling step size. Additionally, the speckle images are utilized to generate many randomly deformed speckle datasets to retrain the model to enhance the generalization of the model in speckle deformation measurement. Finally, the GMA network is combined with the ICGN algorithm, and the performance of the algorithm is evaluated by simulated speckle deformation experiments and real wood block compression experiments.Results and Discussions After optimizing the samp...