LGRF-Net: A Novel Hybrid Attention Network for Lightweight Global Road Feature Extraction

• Published in: IEEE transactions on geoscience and remote sensing 2024, Vol.62, p.1-15
• Main Authors: Duan, Yifei, Qu, Junsuo, Zhang, Le, Qu, Xiaochen, Yang, Dan
• Format: Article
• Language: English
• Subjects: Attention mechanisms; Computational modeling; Convolution; Data mining; Feature extraction; Filters; Global context feature extraction (GCFE); loss function; multiscale feature fusion (MSFF); network efficiency enhancement; Remote sensing; remote sensing image; Roads; Telecommunications; Training
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3491758

In scenarios where road obstacles complicate feature extraction, designing a lightweight convolutional neural network (CNN) model with minimal parameters and flops while maintaining competitive segmentation accuracy poses one of the most challenging research tasks in remote sensing imaging. Finding the optimal balance between segmentation performance and computational efficiency is crucial. We introduce a novel method for global road feature extraction by strategically employing the light ghost basic block to develop a tiny-ghost link network (TG-LinkNet). A multiscale feature fusion (MSFF) module, which combines the parallel channel position attention mechanism (PCPAM) to deliver accurate road structure information, further supports the goal. We present a solution to the issue of feature fusion information retrieval-induced excessive redundant noise, which might cause serious interference. Furthermore, to efficiently extract edge features and capture long-distance reliance on global features, we create a global context feature extraction (GCFE) module, ultimately resulting in the lightweight global road feature extraction network (LGRF-Net). To facilitate efficient training, we implement a 1:2 weight design within our deep supervision technique, termed hybrid loss (weighted cross entropy (WCE)-Dice). Extensive experiments were conducted on the DeepGlobe ( 1024\times 1024 , 512\times 512 ) and SpaceNet road datasets. This demonstrates that our network possesses smaller parameters and flops compared to other road-based semantic segmentation methods.