GAMSNet: Globally aware road detection network with multi-scale residual learning

Road detection from very high-resolution (VHR) remote sensing imagery is of great importance in a broad array of applications. However, the most advanced deep learning based methods often produce fragmented road segments, due to the complex backgrounds of the images, such as the occlusions and shadows caused by trees and buildings, or the surrounding objects with similar textures. In this research, the characteristics of the existing deep learning based road detection methods are analyzed and effective road detection methods are explored, and we show that capturing long-range dependencies can significantly improve the road recognition performance. The novel globally aware road detection network with multi-scale residual learning (GAMS-Net) is proposed, in which multi-scale residual learning is applied to obtain multi-scale features and expand the receptive field, and the global awareness operation is used to capture the spatial context dependencies and inter-channel dependencies. Through capturing useful information over long distances, GAMS-Net can significantly improve the road recognition performance. The advantages of the proposed approach are validated using the public Deep-Globe road dataset and large-scale images, and the experimental results confirm the superiority of the proposed method.