基于非对称卷积网络的遥感影像地物检测方法研究  

Research on Remote Sensing Image Ground Object Detection Method Based on Asymmetric Convolutional Network

作　　者：刘海军 杨鸿海 LIU Haijun;YANG Honghai(Geographic Space and Natural Resources Big Data Center of Qinghai Province,Xining 810001,China;Key Laboratory of Geospatial Information Technology and Application of Qinghai Province,Xining 810001,China)

机构地区：[1]青海省地理空间和自然资源大数据中心,青海西宁810001 [2]青海省地理空间信息技术和应用重点实验室,青海西宁810001

出　　处：《地理空间信息》2023年第9期73-76,共4页Geospatial Information

基　　金：青海省自然科学基金资助项目(2020-ZJ-927)。

摘　　要：针对高分遥感影像地物目标检测时存在的精度低、鲁棒性较差等问题,提出了一种单阶段目标检测模型。利用包含非对称卷积核的AC模块及其等效常规卷积核构建特征提取端分别进行模型训练和测试,使充分提取特征时不会产生大量计算参数;构建4层特征增强端,通过拼接不同倍数的特征图进一步丰富语义特征,利用K-means++聚类获取的锚点框来提高训练初期拟合速度。以RSOD数据集为基础,选取包含飞机、油罐、操场的影像作为数据集,对训练集进行影像质量增强与噪声样本扩充。结果表明,模型在3类目标上的精度表现均衡,综合检测精度达到91.03%;在测试环境下的检测速度可达27 m·s^(-1),达到实时检测的水平。Aiming at the problems of low measurement accuracy and poor robustness in high-resolution remote sensing image ground object detection,we proposed a single-stage target detection model.The model uses the AC module including the asymmetric convolutional kernel and its equivalent conventional convolutional kernel to construct the feature extraction end to train and test the model respectively,to ensure that the model can fully extract features without generating a large number of computational parameters.This model constructs a four-layer feature enhancement end to further enrich semantic features by splicing feature maps of different multiples,and use the anchor boxes obtained by K-means++clustering to improve the fitting speed in the early training stage.Based on the RSOD data set,we selected images including aircraft,oil tanks and playgrounds as the data set,and performed image quality enhancement and noise sample expansion on the training set.The experimental results show that the accuracy of model is balanced on the three types of targets,and the comprehensive detection accuracy reaches 91.03%.The detection speed can reach 27 m·s-1,reaching the level of real-time detection.

关 键 词：高分遥感影像 地物检测 全卷积网络 非对称卷积 多尺度特征增强 

分 类 号：P237[天文地球—摄影测量与遥感]