Please use this identifier to cite or link to this item:
|Title:||Towards high-speed optical measurement of shape and deformation||Authors:||Wang, Tianyi||Keywords:||DRNTU::Engineering::Electrical and electronic engineering::Optics, optoelectronics, photonics
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
DRNTU::Engineering::Computer science and engineering::Software::Programming techniques
|Issue Date:||2018||Source:||Wang, T. (2018). Towards high-speed optical measurement of shape and deformation. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||Optical measurement techniques, with the advantages of non-contact, non-destructiveness, and high accuracy and sensitivity, have been successfully developed and applied to study shapes and deformations of objects. Digital image correlation (DIC) and fringe pattern analysis (FPA) are two important categories of the optical measurement methods. DIC experiments can be performed in almost any imaging technology, while white-light optics are predominantly used. The displacement or strain field is measured by correlation analysis on a random speckle pattern on the surface of an object. On the contrary, FPA is based on regular fringe patterns that are generated by the well-known interference phenomenon. The shape of an object is obtained by phase extraction and phase unwrapping processes. In the past decade, both DIC and FPA had made great progress. However, two key issues still need further improvement. First, in pursuit of higher resolutions for higher measurement accuracy, their computation burden has dramatically increased, making them hardly applied to study dynamic phenomena or integrated into real-time systems. Second, decorrelation is a serious issue when the deformation between the initial and the deformed configurations of an object is large. In fact, these two issues can be remedied by one single solution: increasing the computation efficiency, so that the speed requirement in dynamic experiments can be fulfilled. Large deformation can also be resolved by a high processing speed such that it is possible to insert multiple intermediate frames to make the inter-frame deformation small. In recent years, heterogeneous parallel computing platforms composing of CPUs and Graphics Processing Units (GPUs) have been widely used to increase the computation efficiency of optical measurement techniques due to their cost-effectiveness, easy programming interface, and great portability and scalability, yet details of the employed parallel computing strategies have not been well explained. Therefore, in this thesis, a systematic study of powering optical measurement techniques, especially the DIC and FPA techniques, using CPU and GPU parallel computing are proposed. First, the CPU and GPU parallel computing strategies applied to optical measurement methods, including digital image/volume correlation, fringe pattern analysis, tomography, hyperspectral imaging, computer-generated holograms (CGH), and integral imaging in the past five years are reviewed. It is found that high parallelism with the common four parallel patterns can be always exploited from these methods. Second, a GPU powered high-accuracy parallel DIC (paDIC) algorithm is then proposed. It utilizes a path-independent initial guess transferring scheme so that the calculation at one point of interest (POI) is independent from the others. A 57.5+ times speedup compared with the sequential implementation of the same algorithm has been achieved. It is the fastest DIC algorithm using the 1st order (more accurate than the 0th order) shape function reported in the literature Third, the paDIC is extended to 3D digital volume correlation (DVC) to study the internal deformation of an object, and a parallel DVC (paDVC) is proposed. Due to the introduction of an extra dimension, paDVC is more complex than paDIC. Also, to efficiently process a large amount of 3D data volumes, a batch processing scheme is proposed and employed. The proposed paDVC algorithm is 23.3 times and 3.7 times faster than its sequential and CPU multi-core implementations, respectively. To our best knowledge, this is the fastest DVC algorithm reported so far yet without losing accuracy. Fourth, based on the above two works, a real-time DIC system framework, which combines the strength of both the CPU and the GPU is proposed. While graphical user interface (GUI), data acquisition, processing, and display are pipelined on the CPU, the computationally intensive tasks are off-loaded to the GPU. With the flexibly designed three variations of the framework, a real-time frame rate ranging from 30fps (frames per second) to 130fps has been achieved. Also, a reference updating scheme is employed to compensate the decorrelation when the deformation becomes large. Finally, a real-time reference-based dynamic phase retrieval algorithm called G-LS3U is proposed to extract phase distributions from fringe or speckle patterns. Different parallel computing strategies are developed and applied to both the least-squares fitting and the windowed Fourier filtering (WFF) processes. GLS3U achieved a remarkably high processing rate at 131+ fps, making G-LS3U the fastest reference-based dynamic phase retrieval algorithm reported heretofore.||URI:||http://hdl.handle.net/10356/73234||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.