【作者】 高颖；

【导师】 李林； 刘辉；

【作者基本信息】 燕山大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 随着星载遥感技术的不断提高,高分辨率遥感数据迅猛增加。海量的星载遥感数据对卫星存储能力和下行传输能力提出更高要求,目前有限的卫星数传带宽已无法满足遥感数据的传输需求。该文面向星载舰船检测技术,设计实现基于现场可编程逻辑阵列(Field Programmable Gate Array,FPGA)的光学遥感数据舰船检测系统。首先,根据本文项目来源进行需求分析。按照功能需求,将舰船检测系统划分为板卡设计模块、舰船检测算法硬件实现模块和模拟源软件开发模块。板卡设计模块主要是围绕主控芯片设计硬件电路,为舰船检测系统搭建硬件平台。舰船检测算法硬件实现模块主要用于在嵌入式硬件平台中对遥感数据进行舰船检测,区分海量遥感数据中是否包含舰船目标数据,剔除不包含舰船目标的无效数据,为后级存储设备和传输设备减轻数据传输压力。模拟源软件开发模块主要用于模拟星上运行环境,为舰船检测系统提供数据输入。其次,针对星载应用要求,选择计算密集型、低功耗的FPGA作为舰船检测算法运行的硬件平台。在现有舰船检测算法中选择轻量化、运行速率快的YOLOv3(You Only Look Once)深度学习网络作为舰船检测算法。并针对FPGA硬件特性,在硬件资源有限的条件下采用网络轻量化设计和网络加速化设计,实现YOLOv3深度学习网络在FPGA硬件平台的算法移植。最后,将本文设计的光学遥感数据舰船检测系统在200张可见光遥感数据中进行测试,达到了0.33fps检测速率、85.2%精确率、14.7%虚警率和6.408W运行功耗的实验结果。将本文实验结果与Jetson TX1开发板中的实验结果相比较,单核运算能力提升56倍。本文研究的重大突破是将人工智能常用图形处理器(Graphic Processing Unit,GPU)转变为设计灵活、高并行度的FPGA硬件平台,为后续在国产化FPGA芯片中实现舰船检测系统提供先验设计。更多还原

【Abstract】 With the continuous improvement of satellite-borne remote sensing technology,highresolution remote sensing data has increased rapidly.The massive amount of satellite-borne remote sensing data places higher requirements on satellite storage and downlink transmission capabilities.The current limited satellite data transmission bandwidth can no longer meet the transmission requirements of remote sensing data.This article is oriented to spaceborne ship inspection technology,and designs and implements a field programmable logic array(Field Programmable Gate Array,FPGA)-based optical remote sensing data ship inspection system.First,conduct a demand analysis based on the source of the project in this article.According to functional requirements,the ship inspection system is divided into a board design module,a ship inspection algorithm hardware implementation module,and a simulation source software development module.The board design module is mainly to design the hardware circuit around the main control chip to build a hardware platform for the ship inspection system.The ship detection algorithm hardware implementation module is mainly used for ship detection of remote sensing data in the embedded hardware platform,distinguishing whether the massive remote sensing data contains ship target data,eliminating invalid data that does not contain ship targets,and storing it for the later stage Equipment and transmission equipment reduce the pressure of data transmission.The simulation source software development module is mainly used to simulate the on-board operating environment and provide data input for the ship inspection system.Secondly,in view of the requirements of on-board applications,a computationally intensive,low-power FPGA is selected as the hardware platform for the operation of the ship detection algorithm.In the existing ship detection algorithms,the lightweight and fast running speed YOLOv3(You Only Look Once)deep learning network is selected as the ship detection algorithm.In view of FPGA hardware characteristics,under the condition of limited hardware resources,network lightweight design and network acceleration design are adopted to realize the algorithm migration of YOLOv3 deep learning network on FPGA hardware platform.Finally,the optical remote sensing data ship detection system designed in this paper was tested on 200 pieces of visible light remote sensing data,and it reached the experimental results of 0.33 fps detection rate,85.2% accuracy rate,14.7% false alarm rate and 6.408 W operating power consumption.Comparing the experimental results in this article with the experimental results in the Jetson TX1 development board,the single-core computing power is increased by 56 times.The major breakthrough in this paper is to transform the commonly used graphics processing unit(GPU)for artificial intelligence into an FPGA hardware platform with flexible design and high parallelism,which provides a priori design for the subsequent implementation of the ship inspection system in the localized FPGA chip.更多还原