Research On Method And Realization Of 3D Cement Image Registration Based On GPU

Research on the structure characteristics and formation process of cement pastes is very important in cement materials field, the performance index of cement composites is directly affected by cement hydration reaction and its microstructure. Because of its interior mechanism and uncertainties became more complicated with time; the research of cement hydration faces many challenges. In cement images which are got by traditional methods, the spatial position and angle of cement particles at the same slice of different time are not aligned, so the process of cement hydration evolution can't be observed dynamicly and can't be seen clearly from the same angle at different time.In recent years, image processing is foucusing on image registration. Combine advanced information simulation and traditional experimental method, by using intelligent computation, 3D image and graphics processing technology to reseach microstructure and properties of cement hydration, computer technology will provide new scientific, practicable and efficient methods to conduct the research of cement microstructure. It will have great theoretical and practical significance on cement hydration.In this paper, firstly, cement pastes images of 10 time points in 28 days are scanned using microtomography (μCT), and then grayscale calibration has been achieved to eliminate grayscale difference of the same material; Secondly, a Powell method (PM) and similarity transformation based approach is proposed to make the same cement particles in different time points aligned for three dimensional registration of cement microstructure. Finally, based on GPU CUDA (Compute Unified Device Architecture), this paper proposed a parallel calculation framework for three dimensional image registrations to accelerate image registration speed. The main contents are organized as follows:This paper studies the data acquisition method of cement pastes image by usingμCT. WhenμCT scans the same cement paste at different time, the grayscale of the same matter are different because of the different linear attenuation coefficient (LAC) caused by unstable X-ray's emitted energy. According to cement hydration principle and image processing theory, particle swarm optimization (PSO) is used to optimize the interval parameters'range. The grayscale is firstly redirected to an interval, and then stretche the original gray histogram figure until two wave crests between floating image histogram and reference image histogram aligned, and finally the grayscale correction can be realized. The process of grayscale calibration lays a solid foundation for the microstructure research of cement pastes.In order to make the cement image aligned in space. A Powell method (PM) and similarity transformation based approach is proposed for three dimensional registration of cement microstructure. 3D registration has been implemented for cement pastes images by using microtomography (μCT) in 10 time points of 28 days. Because ofμCT image's good 3D visuality, non-destructive to cement sample and high precision, the microstructure of cement pastes can be observed dynamicly, makes hydration process of cement paste more comprehensive and clear. The experimental result shows that this method is easy to be realized with high efficiency and accuracy, it could be seen as an effective automated registration way.Because of strong interior parallel computing ability exists in the 3D image registration. In order to accelerate the speed of 3D registration, the design and realization of parallel computing architecture based on GPU CUDA for three dimensional image registration is proposed. The procedure of the image registration should be break down firstly, and then the steps which include algorithm optimization, space transformation and interpolation that have the features of large computation and high parallelism were parallel computed by GPU using multithread. The other part which has the features of small computation is still finished by CPU. The experimental result shows that the algorithm performance has been improved effectively, the speedup factor is very conspicuous relative to CPU and the execution time reduces greatly.