| With the development of imaging spectroscopy and the improvement of spectral resolution,the processing technology of hyperspectral remote sensing has received remarkable attention.Hyperspectral remote sensing can detect many small differences in ground objects that are difficult to detect in traditional images or multi-spectral images because of its rich spectral information,enabling researchers to analyze the features more accurate.This makes the hyperspectral target detection shows a unique advantage in solving the detection problem,and has important theoretical significance and practical application value in the field of target detection.However,due to the development of remote sensing methods and imaging technology,the amount of data processed by hyperspectral remote sensing has become larger and larger,which greatly increases the difficulty of data processing.Especially in applications with high-speed requirements,data processing lags seriously.Therefore,how to speed up the processing of data,improve the efficiency of algorithm execution,and improve the timeliness of target detection has become an urgent problem.In recent years,GPU general-purpose computing has become an effective means for fast parallel target detection of hyperspectral remote sensing data with its powerful computing performance and high-speed bandwidth.Aiming at the efficiency problem of hyperspectral remote sensing target detection algorithm,the GPU-based parallel processing method for hyperspectral remote sensing target detection is proposed.This thesis also designs a fast detection method of multi-GPU under multi-thread scheduling,and the effectiveness and efficiency of the algorithm are proved by comparison experiments.Specific work and research results include the following aspects:(1)a GPU parallel optimization method based on storage organization and computing resource tuning is proposed aiming at the hyperspectral video sequence target detection method based on the low-rank decomposition.According to the characteristics of the algorithm,the low-rank matrix decomposition process is optimized by task decomposition and access optimizations of torage organization structure.It also adopts the interaction optimization technology of the global memory and shared memory.Since the matrix dimension involved in the convolution solution is large and the computational complexity of the problem is high,this thesis first analyzes the parallesm of calculation and data,and then reconstructs the parallelization of the algorithm iterative process.The optimization model breaks through the bottleneck of the algorithm performance,effectively reduces the interation time between the CPU and GPU.The real hyperspectral data is used to verify the effectiveness and efficiency of the above optimization method.(2)a low-rank decomposition target detection model based on spatio-temporal TV regularization term is designed on the hybrid heterogeneous platform of CPU+GPU.On the basis of analyzing the parallelism of the algorithm model in detail,the GPU parallel method based on fine-grained thread level is optimized.Combined with the use of computing resources,a method of adaptively configuring the Kernel function dimension is proposed,which makes full use of the existing resources of the computer,and effectively improves the throughput of the GPU.In addition,it solves the problem of multiple iterations of the Lagrangian equation.According to the dependency of the data,reasonable task allocation and storage optimization are carried out,which comprehensively utilize the logic control capability of the CPU and the powerful computing performance of the GPU.Under the condition that the detection accuracy of the serial algorithm is completely consistent,the acceleration ratio of the algorithm after parallel optimization reaches 115.26 x.(3)This thesis proposes a fast algorithm for target detection of hyperspectral video sequences based on multi-GPU platform.Based on the analysis of algorithm data dependence and computational characteristics,a parallel design method based on multi-thread optimization scheduling and multi-task overlapping execution is proposed.The low-latency transmission design based on asynchronous stream technology and multi-type storage allocation based on task level are implemented.Compared with the single GPU method,the multi-GPU method achieves high acceleration and parallel efficiency. |
