During the research of hyper-spectral imaging spectrometer, how to process the huge amount of image data is a difficult problem for all researchers. The amount of image data is about the order of magnitude of several hundred megabytes per second. The only way to solve this problem is parallel computing technology. With the development of multi-core CPU and GPU，parallel computing on multi-core CPU or GPU is increasingly applied in large-scale data processing. In this paper, we propose a new parallel computing solution of hyper-spectral data processing which is based on the multi-CPU and multi-GPU heterogeneous computing platform. We use OpenMP technology to control multi-core CPU, we also use CUDA to schedule the parallel computing on multi-GPU. Experimental results show that the speed of hyper-spectral data processing on the multi-CPU and multi-GPU heterogeneous computing platform is apparently faster than the traditional serial algorithm which is run on single core CPU. Our research has significant meaning for the engineering application of the windowing Fourier transform imaging spectrometer.

Accelerating hyper-spectral data processing on the multi-CPU and multi-GPU heterogeneous computing platform

During the research of hyper-spectral imaging spectrometer, how to process the huge amount of image data is a difficult problem for all researchers The amount of image data is about the order of magnitude of several hundred megabytes per second Traditional solution of the embedded hyper-spectral data processing platform such as DSP and FPGA has its own drawback With the development of GPU, parallel computing on GPU is increasingly applied in large-scale data processing In this paper, we propose a new embedded solution of hyper-spectral data processing platform which is based on the embedded GPU computer We also give a detailed discussion of how to acquire and process hyper-spectral data in embedded GPU computer We use C++ AMP technology to control GPU and schedule the parallel computing Experimental results show that the speed of hyper-spectral data processing on embedded GPU computer is apparently faster than ordinary computer Our research has significant meaning for the engineering application of hyper-spectral imaging spectrometer

A new embedded solution of hyperspectral data processing platform: the embedded GPU computer

Imaging spectroscopy technology combines imaging technology with spectroscopy technology. The most common imaging spectrometers are dispersive imaging spectrometers and interferometric imaging spectrometers. The incident slit of dispersive imaging spectrometers is located on the front focal plane of the collimation system, and the incident light passes through the collimator. After the straight optical system is collimated, the light energy is imaged on different positions of the detector in the order of wavelength by the imaging system after dispersion by the prism or grating, and the target spectrum is directly obtained. The principle is simple and the technology is relatively mature, but the light flux will be affected by the incident slit. The interferometric imaging spectrometer forms an interferogram by introducing different optical path differences into the incident beam, and then uses the Fourier transform relationship between the interferogram and the spectrogram to obtain the spectral information of the object, with high luminous flux, multi-channel and high spectral resolution. The CUDA architecture was developed by NVIDIA and is mainly used in the field of parallel computing. Its core idea is to use the GPU as a co-processor of the CPU, which is responsible for the calculation of a large amount of data and greatly shortens the data calculation time. It has been widely used in finance. Simulation, biological computing, physical simulation, simulation computing and artificial intelligence and other fields. The spectral data processing of the infrared interference spectrometer system mainly includes steps of data rearrangement, de-DC, apodization and Fourier transform. Due to the large amount of calculation of Fourier transform and long calculation time, it seriously affects the real-time application capability of the system. Therefore, this paper proposes a method using base 4 FFT algorithm for implementing fast Fourier transform on GPU platform using CUDA in infrared interference spectrometer system. GPU and the base 4 FFT algorithm are used at the same time to greatly improve the degree of parallelism and shorten the calculation time. A spectral raw image of 640*480*256 band, using the traditional base 2 fast Fourier transform on the CPU takes about 20 minutes to invert into a pseudo-color image. With the algorithm of this paper, the time is shortened to less than 2 minutes.

Research of base 4 FFT algorithm in infrared interference imaging spectrometer based on GPU

Based on the analysis of the classical Zoom-FFT,a Zoom-FFT algorithm based on Fourier transform spectroscopy was presentedBoth FFT and Zoom-FFT were used to process the interferogram with different sampling steps by Matlab simulation software,through the comparison of the inverted and original spectrum curves,it indicated that the spectrum could be inverted by both FFT and Zoom-FFT when the sampling step was less than 20 μm;But when the sampling step was greater than 20 μm and less than 333 μm,the spectrum could not be inverted by FFT algorithm,while the spectrum could be inverted by Zoom-FFT

LI Jun-na

Papers

Zoom-FFT based on Fourier transform spectroscopy