BFNet: a Bi-Frequency Fusion Semantic Segmentation Network for High-Resolution Remote Sensing Images

EasyChair Preprint 14545

Abstract

Semantic segmentation of high-resolution remote sensing images plays an important role in urban planning, traffic guidance and other fields. However, high-resolution remote sensing images usually consist of large and complex scenes and heterogeneous objects, leading to poor segmentation at the edges of objects, which in turn leads to undesirable segmentation of the whole image. To address these problems, we propose an effective bi-frequency fusion semantic segmentation network (BFNet) for high-resolution remote sensing images. Specifically, existing semantic segmentation methods reveal the dominance of CNNs in preserving local ground details, but they still can't globally build when processing full-geomorphic images. To address this problem, BFNet uses a two-branch structure, where the low-frequency branch captures low-frequency context information at different scales based on ESwin-Transformer; meanwhile, a pixel-attention mechanism is designed behind the low-frequency branch to select the optimal global context information; The high-frequency branch extracts high-frequency edge information based on stacked CNNs and transverse connections. In addition, to address the problem of detail loss due to direct fusion of high-frequency and low-frequency information, we designed a boundary fusion module for bi-frequency balancing to enable better segmentation of high-resolution remote sensing images. Our method achieves good performance on two recognized remote sensing datasets, Potsdam and LoveDA, with mIoU of 87.22% on Potsdam and 92.85% on F1. mIoU on LoveDA is 51.37%, which is a relatively good balance in inference speed and accuracy.

Keyphrases: Bi-Frequency Fusion, Boundary Fusion, Pixel Attention, semantic segmentation

@booklet{EasyChair:14545,
  author    = {Chengkun Diao and Jinyu Shi},
  title     = {BFNet: a Bi-Frequency Fusion Semantic Segmentation Network for High-Resolution Remote Sensing Images},
  howpublished = {EasyChair Preprint 14545},
  year      = {EasyChair, 2024}}