利用卷积神经网络进行“问题地图”智能检测  

Intelligent Detection of“Problematic Map”Using Convolutional Neural Network

作　　者：任加新 刘万增[1] 李志林 李然[1] 翟曦 REN Jiaxin;LIU Wanzeng;LI Zhilin;LI Ran;ZHAI Xi(National Geomatics Center of China,Beijing 100830,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China;School of Geosciences and Info-Physics,Central South University,Changsha 410083,China)

机构地区：[1]国家基础地理信息中心,北京100830 [2]西南交通大学地球科学与环境工程学院,四川成都611756 [3]中南大学地球科学与信息物理学院,湖南长沙410083

出　　处：《武汉大学学报（信息科学版）》2021年第4期570-577,共8页Geomatics and Information Science of Wuhan University

基　　金：国家重点研发计划(2018YFC0807005);中国工程科技知识中心建设项目(CKCEST⁃2019⁃2⁃1)。

摘　　要：针对当前中国“问题地图”审核依赖人工目视判别效率低下的问题,提出一种端到端的小样本场景下基于卷积神经网络的多尺度特征融合自适应“问题地图”检测方法。通过对数据集进行实时增强,克服了卷积神经网络需要大量训练样本的问题。通过融合多个不同尺度下的地图,实现了多尺度下的“问题地图”显著错误区域的智能检测。利用版图错误区域属性对区域建议网络进行优化,进一步提高检测的精度。并通过实验验证了所提方法的有效性。相较于现有的“问题地图”检测方法,所提方法的准确率提高8倍,为大规模“问题地图”检测提供了新方法。Objectives:In order to solve the problem of low efficiency of manual visual discrimination in the audit of“problematic map”in China,an end⁃to⁃end adaptive detection method for small sample scene based on the multi⁃scale feature fusion of convolutional neural network(CNN)is proposed in this paper.Methods:The real⁃time enhancement of the dataset can overcome the shortcoming of CNN,which re⁃quires a large number of training samples.By fusing multiple maps at different scales,the intelligent detec⁃tion of significant error areas of the“problematic map”in multiple scales is realized.The region proposal network is optimized by taking the attributes of wrong regions into account to further improve the detection accuracy.Results:Compared with the existing detection method,the accuracy of the proposed method is in⁃creased by 8 times,which verifies its effectiveness.Conclusions:The proposed method provides a new solu⁃tion for large⁃scale“problematic map”detection,and can be quickly applied in production.

关 键 词：“:问题地图” 卷积神经网络 目标检测 多尺度特征融合 小样本场景 

分 类 号：P289[天文地球—地图制图学与地理信息工程]