A Novel Framework for Building Extraction with Dual-path Convolution Network

Building extraction from aerial images has always been a hot research topic in the field of remote sensing image analysis, which is benefit to a wide range of areas. In this letter, a novel framework, Dual-path Convolution Network (D-CN), is proposed for building extraction, which combines different contextual information. The benefits of this framework are twofold: first, the utilization of different area image will increase the confidence of the prediction area greatly. Second, the small size of the predicted feature map will reduce training time, accelerate convergence, but get better performance. Then, we test various variations of D-CN on a public building data set and compare it with Unet, the proposed method can save much training time but get almost the same or better accuracy.