| With the development of remote sensing,imaging spectroscopy has gradually become one of hot research topics in various fields.The imaging spectrometer achieves the goal of the acquisition of spatial image and spectral information simultaneously,and has been widely used in military,environmental detection,geologic survey and other aspects.The interference imaging spectrometer has the advantages in high throughput,multi-channel and high resolution so that many countries have made great efforts in its research and development.The data cube acquired by interference imaging spectrometer contains spatial image and interference fringes of the target.It is a hot research in interference data processing that how to deal with interferogram and reconstruct it into spectrogram with high accuracy in recent years.In most cases,people are more concerned about the position of spectral peaks in the spectrum.In this situation,there is one or more peaks in the whole spectrum,and the values at other locations are usually small or zero.In the absorption spectra of substances,this is also very common,which is manifested in the existence of one or more valleys in the whole absorption spectrum,and other values are larger.That is,such spectra are sparse.Signals with sparse spectral characteristics are widely used in laser interference,spectral material analysis,spectral calibration and laboratory measurement.At present,the traditional Fourier transform is still used to reconstruct the sparse spectrum.If the sparse characteristic can be fully utilized,the efficiency of the restoring algorithm will be further improved.On the other hand,there are still big-data processing problems in the field of spectral reconstruction.In practice,the interference data obtained by interferometer spectrometer is so huge that there are 100,000 or millions of interference sequences in only one scene.For such a large amount of interference data,the traditional serial reconstruction algorithms can not meet the requirements of real time.In addition,in the field of high-resolution spectral estimation,although the restored spectral resolution has been further improved,the current spectral estimation model is extremely time-consuming to solve.It requires a fast solving method.Based on the above analysis,a spectral reconstruction algorithm based on sparse Fourier transform is proposed to the goal of sparse spectrum reconstruction;a high performance parallel processing algorithm for spectrum reconstruction is proposed,for the parallel optimization of QR decomposition,matrix inversion and phase correction with the multiplication method,and a large number of interferometric data are processed in real time on the GPU;an optimization algorithm of autoregressive model based on group iteration is proposed to estimate model parameters in high resolution spectral reconstruction.In addition,the embedded system transplantation of parallel algorithm is completed in this paper,which makes spectrum reconstruction more convenient and flexible.The research contents of this paper are as follows:1.We proposed a spectral reconstruction algorithm based on sparse Fourier transform to improve the algorithm performance for the interference signal whose spectrum is sparse.The algorithm makes full use of the sparse characteristic of the signal in the frequency domain and maps the large value frequency points into a group of buckets according to the mapping rules.Each large-value point in frequency domain is separated by the buckets and then it would be processed by the Fourier transform.The signal is changed from a long sequence to a short one.We design the corresponding restoration rules to reconstruct the original spectrum.Experimental results show that the running time of the algorithm based on sparse Fourier transform for sparse spectrum reconstruction is 0.524 ms,which is almost half of the time required for Fourier transform.2.In this paper,high-performance parallel processing algorithms of interfegrogram based on GPU technology is designed,in order to process massive interference data more quickly and meet the requirements of real-time.The algorithms mainly includes:(1)in the part of eliminating trend terms,we propose the parallel iterative algorithms for QR decomposition and matrix inversion based on the least square method;(2)in the part of apodization,we realize parallel apodization of interference sequence and apodization function using multi-threads on the GPU;(3)in the part of phase correction,we propose a parallel multiplication method to correct the phase error;(4)in the part of Fourier transform,the parallel fast Fourier transform in the parallel computing library CUFFT is used.For the recursive iterative algorithms,we also design the processing mode based on multi-core and single-core.By dividing tasks reasonably on GPU,the dependency relationship between data is broken,and the performance of recovery algorithm is further improved.For batch processing,we further simplify the process of spectrum reconstruction.In the simulation and experiments,the running time of parallel processing scheme is 0.332 ms,and the performance of parallel processing scheme is improved by nearly 70%,compared with serial processing scheme.In batch processing,the performance of parallel processing algorithms are much higher than that of traditional serial processing algorithms.For a frame of actual collected image,there is about 260,000 interference vectors to be processed on the GPU.The algorithms on the GPU are more than 1000 times faster than that of the serial algorithms on the CPU.3.In the high-resolution spectrum reconstruction,the solution of the algorithms,based on modern spectral estimation technology,is very complex.A parallel optimization algorithm for autoregressive model is proposed based on group iteration.Using GPU technology,we solve the parameters in a parallel iteration way for the autoregressive model.The spectrum constructed by the parallel algorithm retains the characteristics of high resolution,and it is completed in a much faster way.At the same time,the parallel solution is used for batch processing,and an asynchronous processing mechanism of interference data is designed to further optimize the parallel processing mode based on the overlap structure of GPU.The experimental results show that the running time of the model parallel algorithm is about 70% shorter than that of the traditional serial algorithm,and the performance of batch processing in high resolution spectral reconstruction is improved by nearly 10 times.The performance of the parallel algorithm is further improved in asynchronous processing mode by 15% to 20%.4.The parallel algorithms we proposed in this paper are implemented on the embedded system to further expand the application of parallel algorithms in spectral reconstruction,which makes spectrum reconstruction more flexible.Using NVIDIA Jetson TX2 development board,the working frequency of parallel algorithm is determined by setting the working mode.By reasonable allocation of resources such as threads and memory,further division and optimization of tasks,parallel algorithms in the development board run much faster.The running time of parallel algorithm under five working modes is tested,compared with that on the CPU cluster.The results show that the performance of parallel algorithm under working mode 0 is the highest,and the running time of parallel algorithm on the embedded GPU is much lower than that on the CPU cluster.For 1000 sets of the same data,the running time of the optimized parallel processing algorithm is 2.925 ms on the PC and 3.331 ms on the embedded system.The performance of embedded system is similar to that of PC. |
PDF Full Text Request