Parallel Algorithm Design And Implementation Of 2D And 3D Digital Image Correlation Combining Scale-invariant Feature

Digital image correlation is a widely-used non-contacting technique for motion and deformation measurement.It can be divided into two types: digital image correlation(DIC)for 2D surface deformation measurement,and digital volume correlation(DVC)for 3D internal deformation measurement on heterogeneous materials.Measurement accuracy and computation efficiency are important issues that restrict the practical application of digital image correlation.Despite the iterative algorithm can achieve high accuracy to 0.01 pixel,its performance depends heavily on the initial guess accuracy,and it is hard to directly apply on large deformation measurement.On the other hand,the demand for high-speed measurement is becoming more urgent.In DIC applications,real-time measurement is getting more and more attention,which requires the computation of an image frame is completed within tens of milliseconds.In DVC,the calculation amount of a single point of interest(POI)has increased by tens to hundreds of times compared with that in DIC,hence DVC has a tight bottleneck of computation efficiency,especially when dealing with high-resolution volume images.Besides,complicated algorithms introduced to improve measurement accuracy often further exacerbate the bottleneck of efficiency.How to develop high-speed and high-accuracy digital image correlation methods is an urgent problem.To address issues mentioned above,this thesis deeply researched 2D and 3D digital image correlation that combine scale-invariant image feature and parallel computing technique.The main research work is as follows:(1)Applied scale-invariant feature transform(SIFT)on digital image correlation,and proposed SIFT-aided path-independent DIC(Pi DIC)and 3D SIFT-aided path-independent DVC(Pi DVC)parallel algorithms.In the algorithm design,novel initial guess methods are proposed,which utilize local image matched points to assist DIC and DVC in estimating initial parameters.Experiments show that the proposed DIC and DVC algorithms can achieve high accuracy and precision on images with large deformation.(2)Implemented the SIFT Pi DIC parallel algorithm based on multi-core CPU and GPU.Task parallelism model is used,and the computation task of each POI is assigned to a CPU thread or a GPU thread block according to the characteristics of the processors.Experiments on multiple different image sequences show that parallel computing improves the computation speed significantly.Compared with the multi-core CPU implementation(10 cores,20 threads),the GPU implementation achieves a 10.6 times speedup on a graphics card containing 3840 stream processors,and achieves high-accuracy real-time measurement speed for normal images.(3)Implemented the 3D SIFT Pi DVC parallel algorithm based on multi-GPU system.The parallel design and implementation of each steps of the algorithm are carefully developed to exploit multi-GPU.To solve the scalability problem of 3D SIFT,a blocking mechanism without loss of accuracy is developed.Experiments on simulated data and real test data show that the multi-GPU implementation can effectively further improve the computation speed.In the heterogeneous dual GPU experiment platform,the computation speed of single GPU is 2919 and 5357 POI/s respectively,and the dual GPU parallel implementation achieves a speed up of 2.68 times(7822 POI/s)compared to the former.