Integrating saliency and ResNet for airport detection in large-size remote sensing images

Abstract

Automatic airport detection has received great attention due to the importance of airports in both military and civilian uses. This paper focuses on automatic airport detection in large-size remote sensing images under a two-step object detection framework. In the first step, both geometrical saliency and local entropy saliency are improved to find more accurate ROIs for detecting airports in large-size remote sensing images. The geometrical saliency is based on line features of airports, and line segment detector (LSD) is used to detect line segments. Then, line group weighted saliency map is generated after line connection, and local entropy saliency map is created by further considering the entropy difference between neighbor pixels. Finally, ROIs can be obtained by combining these two saliency maps. The improved saliencies could make airports prominent as a whole instead of many separated parts, and it could find more accurate ROIs than traditional saliency methods. In the second step, deep residual learning network (ResNet) is used to determine whether a ROI should be labeled as an airport. The unobvious features of airports could be further extracted by ResNet, which greatly promotes the robustness of the proposed two-step object detection method. The experiments on large-size remote sensing images have shown that the proposed method could reduce false alarms greatly compared with both the traditional two-way saliency (TWS) airport detection method and the state-of-the-art object detection method of single shot multi-box detector (SSD).