Research And Implementation Of FPGA Accelerating Anomaly Detection In Hyperspectral Images

In recent years,the anomaly detection based on hyperspectral images has become a research hotspot in the field of remote sensing image processing.Many new algorithms have been proposed and achieved good detection results.However,due to the huge amount of data in hyperspectral images and the slow processing speed of software platform,many anomaly detection algorithms cannot meet the needs of real-time detection.FPGA is a hardware platform which is reprogrammable,radiation-resistant,low-power consumption and small in size.It is endowed with rich logic,storage,and computing resources,which make it possible for the real-time anomaly detection of satellite borne hyperspectral images.But software algorithms often cannot reach the best results and speed when they are directly transplanted to the FPGA platform.Therefore,it is of great research significance to improve the anomaly detection algorithms and accelerate their implementation based on FPGA.AED is a hyperspectral anomaly detection algorithm based on morphological attribute filtering and domain transform recursive filtering,which has achieved good detection results on many data sets.However,the morphological attribute filtering and domain transform recursive filtering of AED algorithm is poor in real-time performance,and they are not suitable for hardware implementation.Therefore,this thesis proposes a real-time anomaly detection algorithm RT-MED,based on FPGA,to improve AED algorithm.Firstly,morphological opening and closing operations are used to replace attribute filtering module to extract the profile of target in images.The opening and closing operations can be decomposed into a combination of erosion operation and dilation operation and can be processed in pipeline ways on FPGA.Then,guided filtering is used to replace domain transform recursive filtering to carry out post-processing on the image,which can not only keep the edge information,but also filter out the interference in the background,and simplify the guided filtering so that its detection accuracy is ensured and processing time is reduced.Finally,based on Xilinx Vivado and HLS tools,the RT-MED algorithm is accelerated based on FPGA platform,and the real-time anomaly detection system of hyperspectral image is implemented on Xilinx VC709 development kit.In order to prove the advantages of the design proposed in this thesis,we used San Diego and ABU data sets to test the anomaly detection systems.The test results showed that the detection performance of RT-MED algorithm is better than existing algorithms such as RX,LRX,CRD and AED.The AUC’s score is higher,target in the detection image is clearer,background interference is less than others’.In the performance comparison test of software and hardware platform,the speed of algorithms based on FPGA is 3.36 times faster than that based on MATLAB,2.68 times faster than that based on C++,and the power consumption is only 20.75% of CPU,which can meet the requirements of high real-time performance and low power consumption.Therefore,the FPGA-based real-time anomaly detection algorithm for hyperspectral images has high practical value,and also provides a reference for the accelerated implementation of other hyperspectral images classification and detection algorithms.