基于多结构卷积神经网络的高分遥感影像飞机目标检测

发布时间：2018-04-06 00:05

本文选题：目标检测　切入点：高分遥感影像　出处：《计算机工程》2017年01期

【摘要】：传统的遥感影像目标检测方法大多利用人工提取特征,难以用于背景复杂的高分辨率遥感影像。针对该问题,构建一种多结构卷积神经网络模型(MSCNN)自动学习目标特征。通过改变卷积滤波器尺寸、数量以及网络层数,分别设计4种不同结构的CNN以提取目标从低层、中层到高层不同尺度的特征信息,并将4种CNN输出采用串行方式连接并输入到BP神经网络分类器进行训练。在检测阶段采用滑动窗口方法进行目标搜索。对高分辨遥感影像中飞机的检测实验结果表明,MSCNN在虚警率和召回率上较4种单一结构的CNN具有明显的检测优势,召回率平均提升6%,虚警率平均降低3%。对油罐的检测结果进一步表明,MSCNN可以推广到对遥感影像其他目标的检测。
[Abstract]:Traditional methods of target detection in remote sensing images mostly use artificial feature extraction, which is difficult to be used in high-resolution remote sensing images with complicated background.In order to solve this problem, a multi-structure convolution neural network model (MSCNNN) is constructed to automatically learn target features.By changing the size and number of convolution filters and the number of network layers, four kinds of CNN with different structures are designed to extract the feature information of target from lower, middle to higher levels.Four CNN outputs are connected in serial mode and input to BP neural network classifier for training.In the detection phase, the sliding window method is used to search the target.The experimental results of aircraft detection in high-resolution remote sensing images show that MSCNN has obvious advantages in detecting false alarm rate and recall rate. The recall rate is increased by an average of 6 points, and the false alarm rate is reduced by an average of 3%.The results of oil tank detection further show that MSCNN can be extended to detect other targets in remote sensing images.
【作者单位】：上海交通大学自动化系;系统控制与信息处理教育部重点实验室;
【基金】：国家“973”计划项目“高分辨率遥感影像的目标特征描述与数学建模”(2012CB719903) 国家自然科学基金“控制理论与方法——网络系统的设计、控制与优化”(61221003);国家自然科学基金青年科学基金“基于贝叶斯集成的视觉注意模型及其高分辨率影像目标快速检测研究”(41101386)
【分类号】：TP751;TP183

【相似文献】