基于改进YOLOX算法的光学卫星影像车辆目标检测  

Vehicle target detection from optical satellite image based on improved YOLOX algorithm

作　　者：张杰[1] 郭杜杜[1] 娄文 郭凯 ZHANG Jie;GUO Dudu;LOU Wen;GUO Kai(School of Mechanical Engineering,Xinjiang University,Urumuqi 830049,China;Enjoyor Technology Co Ltd,Hangzhou 311400,China)

机构地区：[1]新疆大学机械工程学院,新疆乌鲁木齐830049 [2]银江技术股份有限公司,浙江杭州311400

出　　处：《传感器与微系统》2024年第7期132-135,140,共5页Transducer and Microsystem Technologies

基　　金：自治区自然科学基金资助项目(2019D01C043);浙江省智能交通工程技术研究中心开放项目(2021ERCITZJ—KF05)。

摘　　要：为了改善光学卫星影像中车辆目标检测准确率低、速度慢的问题,提出一种基于改进YOLOX算法的卫星影像车辆目标检测方法。首先,以轻量化模型YOLOX的S版本为基线,采用CSPDarknet—53替换原主干特征提取网络,引入基于卷积块的注意力模块(CBAM)提高特征提取时对车辆目标的关注度。其次,扩充主干特征提取网络输出尺度,并在特征强化提取部分设计了一个双向特征金字塔网络(BFP-net),采用亚像素卷积上采样、横向的跳跃连接和纵向的跨尺度连接实现对不同层级、尺度特征的复用,使得最后输出特征层充分融合了分类和定位信息。实验结果表明:本文算法对large-vehicle和small-vehicle两类车辆的检测准确率分别为88.98%和86.58%,相较于原算法,平均检测准确率提高了5.36%,检测速度达到了58.37fps,具有更好的检测效果。In order to improve the problems of low accuracy and speed of vehicle object detection in optical satellite images,a vehicle target detection method based on improved YOLOX algorithm is proposed.Firstly,taking the S version of the YOLOX model as the baseline,CSPDarknet—53 is used to replace the original backbone feature extraction network,and the convolutional block attention module(CBAM)is introduced to improve the attention to the vehicle target during feature extraction.Then,the output scale of backbone feature extraction network is expanded and a bidirectional feature pyramid network(BFPnet)is designed in feature enhancement extraction part.Sub-pixel convolutional upsampling method,horizontal jump connection and vertical cross-scale connection are used to realize the reuse of different level and scale features,so that the final output feature layer fully integrates the classification and positioning information.The experimental results show that the detection accuracy of the proposed algorithm for large-vehicle and small-vehicle is 88.98%and 86.58%,respectively.Compared with the original algorithm,the average detection accuracy is increased by 5.36%,and the detection speed reaches 58.37 fps,which has a better detection effect.

关 键 词：光学卫星影像 车辆检测 尺度扩充 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]