摘要
由于空间探测任务中需处理的遥感图像数量增多,对目标检测系统的鲁棒性和时效性要求越来越高,将大量遥感数据传输给地面再进行处理已经无法满足空间探测任务的需求.针对此问题,本文开展了基于宇航级NPU的遥感图像在轨目标检测系统研究,以Yolov5s网络为基础,替换与NPU适配度低的部分并引入注意力机制提升精度,解决深度学习算法模型复杂和计算量大导致其在星上难以部署的问题.优化后的网络迭代训练后部署在开发板上,经CPU-NPU并行协同处理,并行执行图像处理三部分,充分利用Yulong810A平台资源.实验显示,优化后的网络部署到星上平台后参数量减少75%,mAP值达71.25%,检测速度为47.67 frame·s^(-1),均超过原版Yolov5s网络,成功构建了一个更轻量快速的目标检测系统.
With the continuous progress and expansion of space exploration missions,the quantity of remote sensing images that need to be processed has been increased substantially.In such a context,the target detection systems are confronted with ever-higher demands in terms of robustness and timeliness.The traditional approach of transmitting a large volume of remote sensing data back to the ground for processing has become infeasible due to various limitations such as communication bandwidth and time delay.To address this critical issue,this research focuses on the on-orbit target detection system of re-mote sensing images,which is based on the astronautics-grade Neural Network Processor(NPU).Specifi-cally,the Yolov5s network is taken as the foundation and optimized.The components with relatively low compatibility with the NPU are replaced,and an attention mechanism is incorporated.This not only overcomes the challenges that the complex network structure and excessive computational requirements of deep learning-based object detection algorithms pose for deployment on satellite processing platforms with limited resources,but also compensates for potential losses in network accuracy.The optimized net-work is trained iteratively on the GPU using the public dataset VOC.After the CPU-NPU parallel co-processing design,the three main aspects of image processing,namely image preprocessing,feature extraction,and target classification and localization,are executed in parallel.This approach maximizes the utilization of the limited computing and storage resources of the Yulong810A platform.Experimen-tal results demonstrate that when the optimized network is deployed on the Yulong810A on-board pro-cessing platform,it achieves remarkable improvements.The number of parameters is significantly re-duced by 75%,and compared with the original Yolov5s network,the accuracy is enhanced.The mean Average Precision(mAP)value reaches 71.25%,and the target detection speed attains 47.67 frames per second(fps),which is more than twice the speed of the original Yolov5s network.In summary,this re-search realizes a more lightweight and faster object detection system,which holds great potential for pro-moting the development and efficiency of space exploration missions.
作者
刘冰
周海
卞春江
成晓蕾
王鹏飞
张彪
LIU Bing;ZHOU Hai;BIAN Chunjiang;CHENG Xiaolei;WANG Pengfei;ZHANG Biao(National Space Science Center,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)
出处
《空间科学学报》
北大核心
2025年第4期1123-1133,共11页
Chinese Journal of Space Science
基金
中国科学院青年创新促进会项目资助(E0293401)。
关键词
空间探测
遥感图像
星上处理平台
深度学习
并行处理设计
Space exploration
Remote sensing images
On-board processing platform
Deep learning
Parallel processing design
作者简介
第一作者:刘冰,男,2000年出生于河南省新乡市,现为中国科学院国家空间科学中心学生,研究方向为智能星上信息处理、星上目标检测系统研究等.E-mail:2492869534@qq.com;通信作者:周海,男,1987年出生于安徽省蚌埠市,现为中国科学院国家空间科学中心智能室副主任,博士生导师,研究员,研究方向为目标探测与识别、遥感图像处理等.E-mail:zhouhai@nssc.ac.cn。
