Skip to content

efficient design for crack segmentation with high-resolution and multi-scale transformer

Notifications You must be signed in to change notification settings

CHDyshli/efficient_crack_transformer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 

Repository files navigation

Efficient Crack Transformer

Efficient Dual-Branch High-Resolution Transformer Design for Crack Segmentation

Abstract

Recent studies show that Transformer-based crack segmentation models deliver impressive results. However, due to the quadratic complexity of the self-attention mechanism, the computational cost becomes prohibitive when handling high-resolution dense prediction tasks. This paper presents an efficient Transformer-based network for crack segmentation, utilizing several specially optimized self-attention modules. The proposed model incorporates a dual-branch structure, consisting of a high-resolution branch that maintains high resolution throughout and a low-resolution branch that progressively reduces resolution. A multi-scale spatial-reduction attention is introduced to the high-resolution branch to capture fine details, while an enhanced ReLU-based linear attention in the low-resolution branch learns global features. A bidirectional cross-resolution feature enhancement module connects the two branches, facilitating comprehensive information fusion. On the combined dataset CrackSeg9k and the scenario-specific datasets Asphalt3k and Concrete3k , the proposed method obtains mIoU of \textbf{81.85%,81.29%, and 86.36%,} respectively, which outperforms the comparative state-of-the-art models and also achieves the best trade-off between efficiency and performance.

Dataset

The code will be released soon.

About

efficient design for crack segmentation with high-resolution and multi-scale transformer

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published