Scholarlink

<p>Semantic segmentation is a foundational task in railway scene perception, with its performance critically influencing the accuracy, robustness, and real-time response capability of downstream applications such as autonomous inspection and safety monitoring. While query-mask-based frameworks, particularly those built on Vision Transformers, have become prevalent for global representation learning, they often struggle in railway environments due to two key limitations: randomly initialized queries that misalign with geometrically structured objects like rails and trackbeds, and the lack of explicit modeling of geometric boundaries in mid-to-high-level features, leading to blurred or fragmented segmentation outputs. To address these challenges, we propose GQ-Former, a novel semantic segmentation network that integrates a feature-guided query module (FGQM) and an edge-enhanced semantic learning module (EESLM). The FGQM leverages low-level features to generate spatially aligned query initializations, mitigating semantic drift and improving decoding consistency. The EESLM enhances mid-to-high-level features through multi-scale edge-aware modeling, incorporating Scharr and Laplacian operators alongside convolutional branches to strengthen boundary perception. Evaluated on RailSem19, GQ-Former achieves 66.74% mIoU while maintaining 19.74 FPS under GPU acceleration. Additional validation on MRSI_Dataset11 further achieves 78.02% mIoU. Ablation studies and visualizations confirm its effectiveness in enhancing feature representation and segmentation stability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GQ-Former: edge-guided query initialization for robust railway scene segmentation

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Jiawei Peng,
  • Liqiang Zhu,
  • Jingyu Hu,
  • Kexun Wang,
  • Zhihao Huang,
  • Baoqing Guo

摘要

Semantic segmentation is a foundational task in railway scene perception, with its performance critically influencing the accuracy, robustness, and real-time response capability of downstream applications such as autonomous inspection and safety monitoring. While query-mask-based frameworks, particularly those built on Vision Transformers, have become prevalent for global representation learning, they often struggle in railway environments due to two key limitations: randomly initialized queries that misalign with geometrically structured objects like rails and trackbeds, and the lack of explicit modeling of geometric boundaries in mid-to-high-level features, leading to blurred or fragmented segmentation outputs. To address these challenges, we propose GQ-Former, a novel semantic segmentation network that integrates a feature-guided query module (FGQM) and an edge-enhanced semantic learning module (EESLM). The FGQM leverages low-level features to generate spatially aligned query initializations, mitigating semantic drift and improving decoding consistency. The EESLM enhances mid-to-high-level features through multi-scale edge-aware modeling, incorporating Scharr and Laplacian operators alongside convolutional branches to strengthen boundary perception. Evaluated on RailSem19, GQ-Former achieves 66.74% mIoU while maintaining 19.74 FPS under GPU acceleration. Additional validation on MRSI_Dataset11 further achieves 78.02% mIoU. Ablation studies and visualizations confirm its effectiveness in enhancing feature representation and segmentation stability.