Multiresolution Airport Detection via Hierarchical Reinforcement Learning Saliency Model

Traditional airport detection methods usually utilize geometric characteristics to locate targets, but they are not suitable for low-resolution remote sensing images. Taking both low and high resolution into account, we present a novel hierarchical reinforcement learning (HRL) saliency model to detect airport target. Different from conventional saliency models focusing on nature images, our HRL model is more effective for multiresolution remote sensing images. According to airport characteristic, we design a reinforcement learning structure to suppress background and highlight interesting airport regions level by level. To generate a final saliency map, we fuse bottom–up region features with top–down line feature based on target attribute, which can restrain other salient regions except for airports. Moreover, a learning stop criterion based on latent Dirichlet allocation (LDA) topic model is proposed at each level to judge the state of saliency detection, thus learning process can be adaptively controlled. Besides, a back-level propagation mechanism is employed to reinforce airport target between levels. HRL saliency model can take the advantage of hierarchical structure to quickly locate interest regions in remote sensing images with large cover area. Furthermore, HRL is robust for illumination and resolution variety. Extensive experimental results on a remote sensing dataset containing 730 images of 40 different airports demonstrate that the proposed HRL model outperforms 18 state-of-the-art saliency models in terms of two popular evaluation measures. Besides, it has significantly higher detection rate than other six airport detection methods.