Dual-Path Multi-Level Feature Fusion Network for Semantic Segmentation of Remote Sensing Image

Different objects with similar spectral features are common in remote sensing images, such as trees and low-vegetation, building and roads. It is important to segment them well for urban planning, traffic navigation, and so on. However, the existing multi-level feature fusion methods ignore the relationship among all features of each level, making these objects hard to distinguish. In this paper, we propose a Dualpath Multi-level Feature Fusion Network (DMFFN) to make good use of the features of backbone. This network includes two paths to fuse the features and model the dependences between them. After getting the features from two paths, we utilize a cross-attention module to decoder them for better segmentation. Experimental results over two datasets show that DMFFN outperforms state-of-the-art methods.