面向星载边缘计算的遥感目标检测算法轻量化优化研究  

Lightweight of Remote Sensing Object Detection Algorithm for Spaceborne Edge Computing

作　　者：张鹏程 武文波[1] 李强[1,2] 曹城华[1] ZHANG Pengcheng;WU Wenbo;LI Qiang;Cao Chenghua(Beijing Institute of Space Mechanics&Electricity,Beijing 100090,China;Northwestern Polytechnical University,Xi’an 710072,China)

机构地区：[1]北京空间机电研究所,北京100090 [2]西北工业大学,西安710072

出　　处：《空间控制技术与应用》2022年第5期86-94,共9页Aerospace Control and Application

基　　金：中国空间技术研究院人才基金项目(WYRC2021HQM)。

摘　　要：遥感图像目标实时检测是遥感领域高效能应用的一项关键技术.基于深度网络的目标检测模型检测率高,但该类算法模型往往参数多、计算量大,导致模型在边缘端设备难以部署,对该类模型进行轻量化优化是当前深度网络目标检测算法应用落地的一个核心问题.本文以YOLOv5s目标检测模型作为基础,将一种卷积核剪枝方法应用于YOLOv5s模型中,先对BN(batch normalization)层缩放因子做稀疏化处理,然后以其为评判标准对卷积核进行剪枝和参数微调,并用遥感飞机数据集进行训练和测试.实验结果证明,该方法可以在裁剪了30~50%的模型参数的情况下,使模型的目标检测性能变化在2%以内,即通过该方法可以有效减少YOLOv5s模型的过拟合,达到降低模型大小的效果.Real-time object detection is a key technology in remote sensing field.The detection rate of object detection models based on deep network is high,but this kind of algorithm model often has lots of parameters and computational load,which makes it difficult to deploy the model on edge devices.Lightweight of this kind of model is a problem for the current deep network object detection algorithm.Based on the YOLOv5s object detection model,a convolution kernel pruning method is applied to the YOLOv5s model,first thinning the BN layer scaling factor,then pruning and fine-tuning the convolution kernel,and training and testing with remote sensing aircraft dataset.The experimental results show that the object detection performance of the model can change within 2%when 30-50%of the model parameters are clipped.The method can effectively reduce the over-fitting of YOLOv5s model and achieve the effect of reducing the size of the model.

关 键 词：遥感图像 目标检测 YOLOv5 剪枝 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]