摘要
针对当前储油罐目标检测研究仅针对独立数据集,缺乏区域性检测的实际应用问题,采用一种基于卷积神经网络的嵌入掩膜分支的多阶段目标检测网络(mask multi-stage network,MMSNet)对天津港储油罐进行检测。该网络以ResNet50和特征金字塔网络(feature pyramid network,FPN)作为特征提取网络,在检测头部分设计并行的多阶段检测分支和掩膜分支。多阶段检测分支通过适配不同IoU的候选框,保证训练阶段和推理阶段拥有相似的分布特性;掩膜分支通过提取不同IoU阈值下的特征图,精确回归目标的边界位置。通过国产GF-2号卫星影像数据,提取了天津港地区的储油罐。实验表明,MMSNet网络在保证召回率为92.6%的情况下,准确率可达96.8%,为港口储油罐的快速识别提供了可能。
For the problems that current researches on oil tank target detection only focus on independent data sets and lack the practical application of regional detection,a mask multi-stage network(MMSNet)based on convolutional neural networks was proposed to detect the oil tank in Tianjin port.The ResNet50 and FPN(feature pyramid network)were taken as feature extraction networks,and parallel multi-stage detection branches and mask branches were designed in the detection head.The multi-stage detection branch ensures that the training stage and reasoning stage have similar distribution characteristics by adapting the candidate boxes of different IoU;the mask branch accurately regresses the boundary position of the target by extracting the characteristic images under different IoU thresholds.The oil storage tanks in the Tianjin port area were extracted from the domestic GF-2 satellite image data.Experiments show that the MMSNet network can ensure the recall rate of 92.6%and the accuracy rate of 96.8%,which provides the possibility for the rapid identification of port oil storage tanks.
作者
郝志航
张小咏
陈正超
卢凯旋
HAO Zhihang;ZHANG Xiaoyong;CHEN Zhengchao;LU Kaixuan(Beijing Key Laboratory of High Dynamic Navigation Technology,Beijing Information Science&Technology University,Beijing 100101,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China)
出处
《北京信息科技大学学报（自然科学版）》
2022年第2期8-14,共7页
Journal of Beijing Information Science and Technology University
基金
国家自然科学基金资助项目(41871348)
国家重点研发国际合作资助项目(2019YFE0127705)。
关键词
深度学习
目标检测
储油罐提取
卷积神经网络
deep learning
target detection
oil storage tank extraction
convolutional neural network
