| Hyperspectral remote sensing image plays a great role in target detection,but the rich spectral information of hyperspectral image will lead to the problem of long processing time.In addition,the spectral information contains redundant data,which will have adverse effects on Hyperspectral target detection.Using band selection to eliminate redundant bands can not only avoid the negative impact of redundant information on the subsequent target detection performance,but also reduce the amount of data and improve the detection efficiency.With the parallel optimization of GPU,the computation and complexity of band selection and target detection algorithm can be reduced,so as to improve the efficiency of target detection.This thesis studies the band selection algorithm and target detection algorithm of hyperspectral image,proposes an improved band selection algorithm and target detection algorithm based on orthogonal subspace projection,and optimizes the two algorithms by GPU.The main work and contributions of this thesis are as follows:Firstly,an improved SQ-BSS band selection algorithm is proposed.Aiming at the problem of insufficient parallelism of the search method in the algorithm,an improvement is made.The inner and outer loops in the search method are exchanged to increase the amount of data that can be processed in parallel in the whole algorithm.Then the kernel function is designed and the algorithm is optimized by using the CUBLAS library.The algorithm is optimized by using multiple CUDA streams and asynchronous transmission.The experimental results on different sizes of hyperspectral images show that the performance of the improved SQ-BSS algorithm is similar to that of SQ-BSS,which can improve the detection accuracy of subsequent detection,but the efficiency of the improved SQ-BSS algorithm is higher after parallelization.Secondly,a target detection method based on orthogonal subspace projection is proposed to solve the problem that the target spectral characteristics are known and the background spectral characteristics are unknown.The algorithm uses ATGP and ISODATA to get cluster centers,and spectral angle matching to exclude possible target spectra of interest to obtain the posterior spectral characteristics of the background.Then the hyperspectral image is projected to the orthogonal complement space of the background spectral characteristics to suppress the background signal.Finally,the signal from the background is suppressed and projected into the subspace of the target spectrum to improve the signal-to-noise ratio.The comparison experiments with LCMV and RXD show that the algorithm improves the detection performance compared with traditional algorithms.Next,the time complexity of the algorithm is analyzed,focusing on the parallel optimization of ATGP algorithm and ISODATA algorithm which take a long time.The experimental results show that the parallel optimization of the algorithm can not only ensure the detection performance,but also significantly improve the detection efficiency.Compared with the parallel processing of ATGP and ISODATA in other literatures,this algorithm has greater efficiency.The improved SQ-BSS band selection algorithm and the target detection algorithm based on orthogonal subspace projection are applied to actual projects to improve the target detection rate.After parallel optimization,under NVIDIA Ge Force RTX 2080 Ti,the running time of a single hyperspectral image of 256x256x224 is less than 2s,which shortens the overall running time of the project. |
