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.%Most traditional object detection approaches extract features manually,which can hardly detect specific objects in complex High Resolution Remote Sensing Imagery (HRRSI).For solving the object detection in HRRSI,a Multi-Structure Convolutional Neural Network (MSCNN) model is constructed to learn object features automatically.Four CNNs with different network structures are designed to extract features from low-level,mid-level to high-level with respect to various scales by considering the size and as well as the number of convolution filter in the convolution layers and the number of layer.Then,the outputs of the four CNNs are concentrated and put into a BP network for training a classifier.This paper uses the sliding-window method to search the object.Experimental result on airplane detection in HRRSI shows that MSCNN has obvious advantages than single-structure CNN.It not only reduces the false alarm rate by 3％,but also improves the recall rate by 6％.Experimental result on oil tank detection further shows that MSCNN can be used in the detection of the other objects in remote sensing imageies.
展开▼
