PLFCN: Pyramid Loss Reinforced Fully Convolutional Network

In the field of remote sensing, the semantic segmentation network for orthophotos has received widely attention. However, it is usually impossible to achieve high accuracy and high efficiency at the same time. In this paper, we propose a novel pyramid loss reinforced fully convolutional network (PLFCN) to address this issue. By introducing deep pyramid supervisions, the network explores multi-scale spatial context information to improve performance of semantic segmentation. And the auxiliary pyramid loss structure can be ignored during testing, so that the network can inference as fast as FCN. The main contributions of this paper are as follows: 1) auxiliary pyramid loss structure is proposed to enhance the performance of FCN by multiscale and deep supervisions; 2) the advantages of multi scale structures and auxiliary loss is combined to improve the performance and maintain the efficiency at the same time. The results show that the semantic segmentation performance is significantly improved, while achieves the high effeciency as FCN.