Noise and shuffle feature perturbation driven robust curve segmentation
Jinghao Liu, Li Chen, Kai Zhu, Yunxiang Cao · Journal of Electronic Imaging · 2025
Curve structure segmentation is a crucial field of semantic segmentation, with a wide range of applications in blood vessel segmentation, crack detection, and other fields. Current mainstream curve structure segmentation models typically adopt an integrated architecture design, and their performance improvement tends to rely on the redesign of the model structure or the increase in complexity. To address this issue, we propose a lightweight feature perturbation mechanism. This mechanism enables flexible integration with existing segmentation models and incurs no additional computational complexity overhead. By introducing noise into the feature space, which is combined with the original features of the input image, the model exhibits improved robustness by preserving essential information in the feature representation. We incorporate shuffle operations into the local self-attention mechanism, offering a way to perturb features. This method breaks the fixed associations between pixels and enhances the diversity of features. Then, it is combined with the Mamba model. As a result, the model can more accurately capture the local details and global context information of the curve structures. The experimental results show that the proposed method has made certain improvements over the existing state-of-the-art methods on multiple standard datasets, providing an effective method for curve structure segmentation tasks.