Parallel Computing Algorithm Design And Heterogeneous Computing Implementation Of High Accuracy Digital Image Correlation

Digital Image Correlation,a novel non-contact full-field deformation measurement,mainly includes Digital Image Correlation(DIC)for surface measurement of material,and Digital Volume Correlation(DVC)for internal measurement of material.Two key issues in the research of digital image correlation are the improvement of measurement accuracy and computation efficiency.For processing plane image digital image correlation method,though some scholars keep improving the computation efficiency of DIC method,the computation efficiency is far failed to meet the real-time analysis of high resolution.For stereoscopic image processing of digital volume correlation method,traditional matching method based on subset is hard to obtain high accuracy when facing complex internal deformation.What's more,the computation efficiency has become an important factor restricting DVC methods because the the amount of data in digital volume correlation method has increased by 2 to 3 orders of magnitude and the calculation process is more complex than the digital image correlation.In order to address these issuses and challenges,this thesis has made an in-depth study of DIC and DVC method.The main work and innovation points are as follows:(1)Proposed the Hybrid DIC algorithm based on CPU-GPU heterogeneous computing model.This thesis combines the task parallel and data parallel mode to design the CPU-GPU heterogenous computing model Hybrid DIC,which can make full use of the computing resources of CPU and GPU.This model can finally achieve the real time processing speed of high resolution in the real material image sequence test,the real time processing speed means that processing 24 frames images in 1 second,each image contains about 10000 Point of Interest,POI).On the other hand,a detailed analysis of the data format to DIC's accuracy is completed,which shows the use of single precision data format can achieve the optimal performance without losing significant accuracy.(2)Proposed the SIFT-DVC algorithm combined with 3D-SIFT algorithm and DVC algorithm.Since the traditional subset-based method is hard to achieve high accuracy when facing with complex deformation,this thesis introduced the 3D-SIFT(Scale Invariant Feature Transform)algorithm to extract the robust image feature,combining with the classical Inverse composition Gauss-Newton algorithm to propose a novel SIFT-DVC algorithm.This method firstly extracts the keypoints in reference volume image and target volume image and match them,then use the matching keypoints to carry out an affine transformation matrix,which can be utilized to acquire the initial guess value.Finally,take the initial guess value as input to ICGN algorithm to get the high accuracy result through iterative processing.Compared with the existing PIDVC algorithm,this method can significantly reduce the mean bias error on the volume image with the rotation of more than 5 degrees.(3)the 3D-SIFT algorithm and IC-GN algorithm are implemented based on GPU.The 3D-SIFT algorithm and IC-GN algorithm are complex,the computation is large,thus the computation speed is slow.In this thesis,each computation step of both the 3D-SIFT algorithm and IC-GN algorithm are analyzed for parallelization in detail,and the heterogeneous computing acceleration of the two algorithms are implemented.For both the 3D-SIFT and SIFT-DVC algorithm,the computation speed of GPU-based implementation has improved significantly compared to CPU-based implementation.