Abstract:

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. The experimental results with a detection rate of 93.6%, a false alarm rate of 11.6%, and the time of 0.2s are superior to the results by other comparison methods.

Key words: airport detection, region-based convolutional neural network, transfer learning, hard example mining, alternating optimization