# DGCFNet: Dual Global Context Fusion Network for remote sensing image semantic segmentation

## TL;DR

This paper introduces DGCFNet, a new network for segmenting remote sensing images by combining CNNs and Transformers to better capture local and global features.

## Contribution

The novel DGCFNet integrates CNN and Transformer strengths with dual-branch global extraction and cross-level interaction modules for improved segmentation.

## Key findings

- DGCFNet achieves mIoU scores of 82.20% on Vaihingen, 83.84% on Potsdam, and 68.87% on BLU datasets.
- The dual-branch global extraction module enhances global context modeling while preserving local details.
- Cross-level information interaction improves feature correlation across different levels for better segmentation.

## Abstract

The semantic segmentation task of remote sensing images often faces various challenges such as complex backgrounds, high inter-class similarity, and significant differences in intra-class visual attributes. Therefore, segmentation models need to capture both rich local information and long-distance contextual information to overcome these challenges. Although convolutional neural networks (CNNs) have strong capabilities in extracting local information, they are limited in establishing long-range dependencies due to the inherent limitations of convolution. While Transformer can extract long-range contextual information through multi-head self attention mechanism, which has significant advantages in capturing global feature dependencies. To achieve high-precision semantic segmentation of remote sensing images, this article proposes a novel remote sensing image semantic segmentation network, named the Dual Global Context Fusion Network (DGCFNet), which is based on an encoder-decoder structure and integrates the advantages of CNN in capturing local information and Transformer in establishing remote contextual information. Specifically, to further enhance the ability of Transformer in modeling global context, a dual-branch global extraction module is proposed, in which the global compensation branch can not only supplement global information but also preserve local information. In addition, to increase the attention to salient regions, a cross-level information interaction module is adopted to enhance the correlation between features at different levels. Finally, to optimize the continuity and consistency of segmentation results, a feature interaction guided module is used to adaptively fuse information from intra layer and inter layer. Extensive experiments on the Vaihingen, Potsdam, and BLU datasets have shown that the proposed DGCFNet method can achieve better segmentation performance, with mIoU reaching 82.20%, 83.84% and 68.87%, respectively.