An Effective Attention-Guided Feature Fusion Network for Segmentation of Remote Sensing Imagery

摘要

Recently, the features learned by FCN-based image segmentation methods are usually not efficient enough to the remote sensing imagery. One major reason is that the loss of detailed features high-level layers in these models. An another reason is that the low-level layers lack semantic information. So we propose feature fusion network for segmentation of remote sensing images. To tackle the relatively problems we introduce our network which contains two structures: enhance network and fusion network. The enhance network devotes to capture more effective low-level features, especially small object and boundary features. And the fusion network devotes to fuse semantic information into low-level features and fuse more detail information into high-level features. Based on our proposed feature fusion network, a large number of experiments show that our proposed method performs well on many datasets, especially in segmentation of small target and boundary.