迁移学习结合难分样本挖掘的机场目标检测

Airport object detection combining transfer learning and hard example mining

为了提高遥感图像中机场检测的准确性和速度,提出一种迁移学习结合难分样本挖掘的机场检测方法.首先,舍弃以往滑动窗口加手工设计特征的方式,构造了区域卷积神经网络作为基本架构;其次,基于自然图像和机场遥感图像具有共同的低级和中级视觉特征,网络模型在自然图像上进行预训练并修改完善后,在数据有限的机场上迁移学习;然后,在样本训练中借鉴难分样本挖掘思想来提高网络的目标判别能力和训练效能;最后,使用交叉优化策略实现区域建议网络和后续检测网络的卷积层共享,大大地减少了检测时间.仿真结果表明,所提方法能在复杂背景下准确地检测出不同类型的机场,得到检测率为93.6%、虚警率为11.6%、时间为0.2s的实验结果,各项性能均优于其他对比方法.

In order to improve the accuracy and speed of airport detection of remote sensing images,an airport detection method combining transfer learning and hard example mining is proposed.First,the region-based convolutional neural network is used as the basic framework instead of the sliding windows plus artificial features in the traditional methods.Second,based on the common low-level and intermediate-level visual features between natural images and remote sensing images,the pre-training network trained on natural images is transferred to deal with airport detection with limited data after modifying and improving the network.Then,the idea of hard example mining is introduced to improve the ability to discriminate objects and make the training more efficient.Finally,the alternating optimization strategy allows for sharing convolution layers between regional proposal network and detection network,thus greatly reducing the time.Experimental results show that the proposed method can detect different types of airports accurately under complex backgrounds.T'he experimental results with a detection rate of 93.6%,a false alarm rate of 11.6%,and the time of 0.2 s are superior to the results by other comparison methods.