提出一种CNN的遥感图像飞机检测的方法。首先获得预训练好的CNN, 然后通过参数迁移获得五层卷积层模型参数, 接着利用遥感图像对第五层卷积层进行微调获得一个特征提取器。将特征提取器用于提取遥感图像训练集的深度特征, 训练可变形部件检测模型。实验表明, 提出的方法大大提高了遥感图像飞机目标检测精度, 准确率达96%以上。

A new algorithm for airplane detection in remote sensing image was proposed based on convolutional neural network. Firstly the pre-trained CNN was prepared. Then model parameters of the first five layers of the model were obtained by transfer learning. In the end, the fifth layer was finetuned by remote sensing image and a model of feature extraction was obtained. The obtained model was used to extract deep feature of train image set of remote sensing image for training deformable part model. Experiment shows that by this algorithm , the accurate rate of localization could reach 96%.