Real-time Processing Of Hyperspectral Image Target Detection Based On Deep Feature Extraction

Hyperspectral images can provide certain morphological information and rich information about the target area in both spatial and spectral dimensions.Therefore,target detection technology based on hyperspectral images plays an important role in camouflage military target detection and natural disaster early warning.The hyperspectral detector on the satellite platform can cover most parts of the earth’s surface,which can monitor the target area more effectively.However,the huge data volume of hyperspectral images puts a lot of pressure on the storage and communication channels of satellite equipment.So,there is an urgent need and high value for real-time target detection and processing of hyperspectral images on spaceborne platforms with limited resources and power consumption.In response to the real-time processing requirements of hyperspectral images,a low complexity hyperspectral object detection algorithm based on spectral regularization is proposed.And the use of high-level synthesis tools and hardware description language combining the algorithm hardware acceleration circuit design on the small size,low power consumption,and resource-rich FPGA platform.On this basis,a real-time processing system for hyperspectral image target detection is proposed.The main research contents and research results can be summarized as follows:Firstly,aiming at the problems of feature extraction module’s strong data dependence,feature selection module’s high computational complexity,and detector’s using the dynamic window in hyperspectral image target detection algorithm based on spectral regularization,this paper optimizes the algorithm.First,the feature extraction network is simplified as one layer full connection layer computation.Second,the complex spectrum angle calculation in the feature selection process is removed,and the approximate cosine value is used for feature map selection.Third,fixed weights are used instead of calculating weights based on tensors to fuse feature images.Fourth,a hardware-friendly rectangular window is used to replace the connected region in the morphological filter.Finally,the computational flow of the guided filter is simplified to further reduce the complexity of the algorithm.Secondly,the circuit design of the feature extraction module and feature fusion module is accelerated by the way of parallel between lines and pipelining within lines.And algorithm parameters and circuit structure can be adjusted conveniently through parameterized design.In this paper,a flexible and reusable sliding window filter circuit is designed by the way of row and column filtering,and a high-speed full pipeline morphological filter and guided filter acceleration circuit are proposed.Lastly,combining the designed high-efficiency data storage interactive circuit and systemlevel optimization design,this paper proposes a real-time processing system for hyperspectral image target detection.According to the application requirements in different scenarios of the satellite platform,the hardware circuit can be flexibly configured for different image sizes.The experimental results on Xilinx xc7vx690t FPGA show that the system can accelerate the hyperspectral image processing with the resource utilization of 19% LUT,9% FF,13% Bram,and 2% DSP,and the clock frequency of 200 MHz.In the test of the ABU data set,the system can process a hyperspectral image with a resolution of 100 × 100 and 224 spectral segments in 0.021 seconds,which meets the high-precision,real-time processing requirements.