一种基于深度学习的光学遥感影像在轨目标检测方法  

An On-Orbit Object Detection Method Based on Deep Learning for Optical Remote Sensing Image

作　　者：璩泽旭[1] 方火能[1] 肖化超[1] 张佳鹏[1] 袁玉 张超[1] QU Zexu;FANG Huoneng;XIAO Huachao;ZHANG Jiapeng;YUAN Yu;ZHANG Chao(Xi’an Institute of Space Radio Technology of Space Technology,Xi’an 710000,China)

机构地区：[1]西安空间无线电技术研究所,西安710000

出　　处：《空间控制技术与应用》2022年第5期105-115,共11页Aerospace Control and Application

基　　金：国家自然科学基金资助项目(62171342)。

摘　　要：针对遥感影像复杂的目标成像特性,采用传统的目标检测算法准确率低、鲁棒性不够的问题,提出了一种基于深度学习的光学遥感影像在轨目标检测方法.在硬件层面,设计了大规模可编程逻辑器件FPGA与多核DSP为构架的星上硬件处理平台,支持在轨目标检测网络参数上注重构功能,实现深度学习模型性能不断优化.在软件层面,采用模块化、参数化和并行流水等设计思想的软件架构和数据流,有效提升了算法实现的效率和可移植性.在算法层面,该方法在YOLOv3特征提取网络(DarkNet-53)的基础上引入深度分离卷积(depthwise separable convolution)以有效压缩模型参数与推理计算量.在检测阶段加入局部再检测模块以提升算法对密集目标的适应性.硬件实测结果表明,与目前常用的目标检测方法相比,该方法在处理速度和精度上都有较大的提升,目标检测精度高于90%,单元处理速度达到334.24FPS.同时支持飞机、舰船、车辆等典型目标的检测,为型号应用奠定基础.Aiming at the complex target imaging characteristics of remote sensing images,the traditional object detection and recognition technology has the problems of low accuracy and insufficient robustness.An on-orbit object detection method is proposed in this paper based on deep learning for optical remote sensing image.At the hardware level,this method uses FPGA and multi-core DSP to build an on-board hardware processing platform,which can update the on-orbit object library and optimize the performance of the deep learning model.At the algorithmic level,this method introduces deep separation convolution based on YOLOv3 feature extraction network(Darknet-53)to effectively compress model parameters and inference computation.In the detection stage,a local re-detection module is added to improve the adaptability of the algorithm to dense targets.Compared with the current object detection methods,this method has a great improvement in processing speed and accuracy,the detection accuracy of the target is higher than 90%,and the unit processing speed reaches 334.24FPS.At the same time,it supports the detection of aircraft,ships,vehicles and other typical targets,laying a foundation for model application.

关 键 词：遥感影像 目标检测 深度学习 YOLOv3 FPGA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]