Research On Hyperspectral Remote Sensing Target Detection Parallel Processing

Spectral imaging technology appeared in the 1980s. It is an innovation that the technology combine the spectral information which material composition determines to spatial characteristic. Later, people apply this technology to non-contact target detection, then extract information, processing and analysis. That is called remote sensing technology. As technology advances, in just a few decades, remote sensing technology has experienced multispectral remote sensing and hyperspectral remote sensing. Compared with the multispectral remote sensing, hyperspectral can provide more abundant spectral information.The spectral resolution is higher. There is a huge amount of data, greatly leading to a difficulty in data processing. For some applications which demand higher real-time processing,there is a serious lag. Therefore,it is an urgent problem that how to accelerate data speed, improve the execution efficiency of the algorithm, improve the efficiency of the target detection.Based on the above problems, this paper studies the efficiency problem of target detection, a real-time target detection algorithm is put forward based on the GPU parallel processing. For low speed of computing coefficient in target matching algorithm, this paper propose an algorithm based on causal dictionary residual updates fast OMP algorithm,and implements the parallel processing on GPU. For the efficiency of KRX algorithm, this paper puts forward a nuclear recursive anomaly target detection algorithm. The main contents include the following aspects:First of all, in practical applications, the fast processing in hyperspectral remote sensing demand low algorithm complexity and high data operation efficiency. The effects of traditional RX algorithm and its related algorithms are not ideal. Computational efficiency is low. This paper shows the real-time anomaly target detection algorithm derivation, and put forward parallel processing method under the GPU architecture. It make full use of the GPU computing core, through CUDA programming to realize the acceleration of the hyperspectral data calculation. It can meet the demand of some applications for real-time processing.Secondly, sparse theory is gradually applied to hyperspectral target detection algorithm.But the speed of computing sparse coefficient is low. In order to solve this problem,this paper proposes a fast OMP algorithm based on causal dictionary residual updates by applying the Hermitian lemma, avoiding the repeated calculation of higher dimensional matrix data. In order to further improve the execution efficiency of algorithm, a large number of threads is used on GPU. Compared with serial algorithm, parallel processing methods achieve up to 33.2 speed-up, realizing fast processing.Finally, KRX algorithm based on nuclear machine learning can make full use of the high nonlinear spectral characteristics between spectral bands. it can get better detection result. But KRX algorithm still has high computational complexity. Therefore, this paper introduce Woodbury lemma to implement Gram matrix update in pixel level, avoiding the repeated computation. The experimental results show that compared with the traditional algorithm,target detection efficiency is improved, meeting the requirements of fast processing.