Improved Deeplabv3 For Better Road Segmentation In Remote Sensing Images
Bo Quan, Biyuan Liu, Daocai Fu, Huaixin Chen, Xiaoyu Liu
Abstract
Road extraction from remote sensing images (RSI) is one of the most important applications in semantic segmentation task. In this paper, we propose an improved DeepLab-V3model for better road segmentation in RSI. An improved Deeplab-V3 network model combined with U-Net fusion shallow features is constructed, and the collective loss function of DICE loss and BCE loss in network feedback learning is used to solve the problem of imbalance of two-class samples and effectively extracts roads in remote sensing scenes. The experiment on a challenging road segmentation dataset from Google Earth confirms that our method is better than that of Deeplab-V3 and U-Net, making Deeplab-V3 more practical for road extraction of RSI.