Real-Time Processing And FPGA Implementation Of Hyperspectral Target Detection And Endmember Extraction

Hyperspectral images have important research significance in target detection and classification of ground objects due to the characteristics of "Combines both spatial information and consecutive spectral bands".However,the huge amount of data of hyperspectral images causes great pressure on real-time processing of algorithms in scenes requiring rapid response,such as satellite and aircraft.With the performance of field programmable gate array(FPGA)greatly improved,the hyperspectral algorithm can be effectively accelerated by virtue of the advantages of high parallel and low power consumption of the hardware platform with FPGA as the core processor to meet the requirements of real-time processing.In this paper,aiming at the hyperspectral endmember extraction and target detection algorithms with high real-time processing requirements,on the basis of the algorithm research,the algorithm is improved according to the characteristics of the FPGA hardware platform,and the real-time processing hardware system scheme and FPGA logic program are designed.Since the Simplex Volume Growth Algorithm(SGA)has been widely studied for its advantages of simple calculation and good performance,this paper selects this algorithm for endmember extraction,analyzes the main difficulties faced by the algorithm in real-time processing,and provides corresponding solutions.The main research contents and results are as follows:First,in order to solve the problem that the SGA is not conducive to hardware platform deployment,this paper implements a low-complexity recursive endmember extraction algorithm ROP-GSGA based on geometric projection.Because the OP-SGA has a matrix inversion operation when calculating the orthogonal projection operator,and as the dimension of the endmember matrix continues to increase,the matrix inversion will become more complex and time-consuming,which causes great difficulties in the design of circuit modules.Therefore,based on the matrix inversion theorem,this paper implements the recursive endmember extraction algorithm ROP-GSGA.The biggest advantage of this algorithm is that it converts the increasingly large and complex matrix inversion operation into a low-complexity calculation with a constant amount of calculation,which is beneficial to the deployment of the algorithm on the hardware platform.At the same time,the corresponding recursive optimization scheme is also given for the similar orthogonal projection problem in the ATGP algorithm.Secondly,in order to meet the processing speed requirement of real-time detection system,the hardware implementation system scheme of the ROP-GSGA based on FPGA is designed,and the specific circuit logic design program of the key core modules is given.In order to further accelerate the execution speed of the algorithm,the inner loop is unrolled and executed in parallel by instantiating multiple multipliers and adders for the matrix multiplication;for the update of the core projection matrix operator,a corresponding recursive inverter is designed,and by inserting the pipeline The register method divides multiple continuous operation steps in the module into simple calculations in a single clock step by step,which can not only achieve continuous data update,but also help improve the stable operating clock frequency of the circuit.In addition,logic multiplexing is used for partial circuits with the same function,and the resource consumption and processing speed are balanced as much as possible under the premise of ensuring real-time requirements.Third,the algorithm implemented by the FPGA logic program is objectively evaluated from the three perspectives of detection accuracy,resource consumption and detection rate by using simulated and actual hyperspectral data respectively.Experiments show that the detection accuracy of the algorithm implemented in this paper is comparable to the original algorithm,and the algorithm execution efficiency is greatly improved compared with software through hardware platform acceleration,and the hardware architecture meets the real-time requirements.In addition,the real-time processing system implemented in this paper meets the requirements of real-time detection while the resource requirements are relatively small,which provides a reference for the hyperspectral image processing project with high spatial resolution.