Design Of Real-time Processing System For Hyperspectral Image Target Detection

The hyperspectral imagery from remote sensing combines both spatial information and hundreds of consecutive spectral bands.By utilizing abundant spectral information,users can accurately detect the object matter of interest.Therefore,when it comes to the space exploration,target detection technology for hyperspectral imagery has received extensive attention and it has been applied to many practical projects.However,there is a contradiction between the massive hyperspectral image data and the limited bandwidth of satellite transmission channel.Objectively speaking,it is required that the real-time detection of interested target must be completed on the satellite,and the parts including detected target of image should be compressed first to reduce the pressure on channel transmission.The key issue is how to achieve on-board real-time processing of hyperspectral image target detection.To this end,this paper focuses on the unsupervised recursive orthogonal subspace projection algorithm(UROSP)with high detection performance.Furthermore,the problems faced by the algorithm in real-time processing are analyzed,and the corresponding solutions are given as well.The specific research work carried out in this paper includes the following three parts.First,this paper presents a real-time unsupervised recursive target detection algorithm(RTUROSP).Aiming at the time-consuming problem caused by orthogonal projection during the undesired object extraction in UROSP,it is proposed that projection via the vector operator can replace the complex matrix projection process.In order to make the hardware implementation easier,given that there are similar operations,it is proposed that similar operations can be shared by one processing unit.Secondly,in order to meet the requirements of real-time target detection system on the satellite for processing speed and resource consumption,the hardware structure based on RT-UROSP algorithm is designed in this paper.In view of the slow speed of serial image projection,the hardware structure designed in this paper consists of multiple parallel processing cores which are used to accelerate the vector projection operation.To achieve the goal of low resource consumption,the core operation module of the hardware structure based on RT-UROSP algorithm can be used both in the undesired target extraction stage and the detection of the expected target function under the scheduling of the detection controlling module.Third,for the purpose of completing the on-board real-time hyperspectral image target detection and reducing the pressure of relay satellite channel transmission bandwidth,a realtime processing system for target detection of hyperspectral image based on Xilinx XC7VX690 T FPGA is implemented.The system includes a RT-UROSP algorithm subsystem capable of detecting target substances in real time,and a fast region of interest compression subsystem based on slope displacement method.The experimental results show that the detection time of the system suppressed 4 undesired targets on the TE1 data set is 7.2ms,which is 44.9 times faster than that of the UROSP algorithm running on MATLAB.The detection time of the system running on the HYDICE Urban data set to suppress 8 undesired targets is 29.8ms,which is 46.5 times faster than that of UROSP algorithm running on MATLAB.The system implemented in this paper consumes a small amount of hardware resources.This hardware system has utilized about 30.99% LUT,56.4% BRAM,and 29.34% DSP48E1 of total on-chip.Therefore,the real-time processing system for hyperspectral target detection realized in this paper satisfies the requirement of high-speed processing and achieves a balance between processing speed and resources.