GPU-Based Implementation Of Terahertz MIMO Array Imaging Algorithm

Terahertz(THz)waves are generally referred to as the electromagnetic waves in the spectral band between 0.1and 10 THz,which lies in the gap between the millimeter wave and infrared.Compared to microwaves,the terahertz waves have shorter wavelengths,which result in much better spatial resolution and precision.In contrast to the X-ray,the THz waves are nonionizing,meaning their photons are not energetic enough to knock electrons off atoms and molecules in human tissue,which could trigger harmful chemical reactions.Unlike visible light and infrared,the waves also stimulate molecular and electronic motions in many materials—reflecting off some,propagating through others,and being absorbed by the rest.These features can be exploited to identify explosives,detect concealed weapons,check for defects in tiles on the space shuttle,screen for skin cancer and tooth decay,security screening,and etc.Typical beamforming methods for the terahertz near-field imaging include “Single-InputSingle-Output(SISO)-scanning” ? “2D(Two dimensional)-MIMO(Multiple-InputMultiple-Output)” ?“MIMO-scanning”,and the like.The “SISO-scanning”-based system is relatively complicated,takes a long time to form an image.Although the “2D-MIMO”-based system has fast imaging speed,its system is much more complicated than that of the “SISO-scanning”-based system.Compared with the SISO array,A MIMO array has higher efficiency of element utilization,element position and low system cost.Therefore,“MIMOscanning” based system could obtain a high frame rate at a lower system complexity,and therefore,the corresponding imaging algorithm becomes an important topic in the research and development(R&D)of the terahertz near-field imaging.The powerful parallel computing capability of the graphic processing unit(GPU)has made it evolve from the original display processor to a general accelerator of computation,which has been widely used in the video processing,intelligent remote sensing,space exploration,and other fields.Based on the background described above,this paper mainly studies the implementation of the GPU-based MIMO array imaging algorithm.Major accomplishments in this effort are summarized as follows:1.Extending the application of the back-projection(BP)imaging algorithm into the imaging system based on the MIMO linear array.The BP algorithm is a popular time-domain imaging algorithm with high imaging accuracy and no requirements on the array configuration.The MIMO time-domain echo model is established and the corresponding BP imaging algorithm is derived.The good focusing performance of the BP algorithm is verified by applying it to a simulated scene with five point-scatters.2.In the terahertz security scanning imaging system,the assumption of the far-field plane wave is no longer available,because the array size,scanning length,and the distance between the observation aperture and the target are close,so that the approximation method of using an equivalent phase center is no longer applicable in this system.For this reason,this paper adopts a fast full-focusing imaging algorithm,which is based on a spherical wave model.Compared with the inefficient BP algorithm by using the point-by-point scanning approach,the fast full-focusing algorithm provides high focusing accuracy and improved efficiency of computation.To compensate the effects of the high-order scattering component and propagation attenuation of the spherical wave that are ignored in the above fast fullfocusing algorithm,we modify the algorithm based on the improved Kirchhoff expression in the frequency domain for the signal model.The simulation results show that the modified algorithm can provide rather good focusing performance.3.Design of the optimized parallel framework of GPU-based MIMO array.Aiming at the problem that the imaging speed is slow due to a large amount of data and the slow scanning speed,which can not meet the real-time requirement on the terahertz security imaging system,a GPU-based parallel optimized framework is designed.For the selected imaging algorithm,a parallel processing scheme based on CPU and GPU is designed to meet the requirements on data-intensive and computational-intensive operations,and a suitable computing platform is selected according to the computation load and data structure.We choose a computing platform consisting of 8 GPUs and 2 CPUs,where GPUs and CPUs perform the algorithm cooperatively and optimally.The consistency of GPU optimization results with those simulated on the MATLAB platform confirms the correctness of the design of the parallel framework.4.The parallel processing of nested loops and the implementation of multi-GPU based parallel algorithm are studied.During the imaging process,data are collected from multiple dimensions.Data structure involves a variety of transformations and a large number of accumulative operations.Because serial processing takes too much time,we convert the serial execution into multi-branch parallel operation based on the logical relationship in the data structure and algorithm.The Open MP and multi-GPU parallel implementation framework are designed with 8 GPUs being working in parallel,which significantly improves the computational efficiency of the algorithm.5.For the possible overflow in the display memory caused by the large amount of data,this paper proposes several solutions including the multi-stream heterogeneous execution and data partitioning.The multi-stream heterogeneous processing approach not only solves the possible overflow in the display memory but also reduces the I/O transmission delay.Because the data access by column FFT transformation is too slow,we also propose a data transformation method to meet the requirement of combined memory access and to improve the operation efficiency.Through the optimization implementation of the above algorithm,a 10-hour task needed by using Matlab is successfully finished in 8 seconds by using the designed framework with 8 GPUs.