CrackSegFlow: Controllable Flow Matching Synthesis for Generalizable Crack Segmentation with a 50K Image-Mask Benchmark

TL;DR

This work tackles data scarcity and domain shift in pixel-level crack segmentation by introducing CrackSegFlow, a controllable Flow Matching framework that jointly learns a mask generator and a mask-conditioned image renderer. The renderer uses topology-preserving mask injection and boundary-gated modulation to maintain sub-pixel crack continuity and reduce false positives, while deterministic ODE-based sampling preserves crack geometry. Across five crack datasets, augmenting real data with CrackSegFlow synthetic pairs yields in-domain gains of about 5.37 mIoU and 5.13 F1 on average, and target-guided cross-domain gains of 13.12 mIoU and 14.82 F1, with large benefits when transferring from challenging sources like CFD. The authors also release CSF-50K, a 50k image–mask benchmark for reproducible evaluation of topology-aware crack synthesis and cross-domain segmentation, and demonstrate that their approach outperforms diffusion-based semantic synthesis in fidelity and efficiency, offering a scalable augmentation strategy for robust infrastructure-crack segmentation.

Abstract

Automated crack segmentation is essential for condition assessment, yet deployment is limited by scarce pixel-level labels and domain shift. We present CrackSegFlow, a controllable flow-matching synthesis framework that generates crack images conditioned on binary masks with mask-image alignment. The renderer combines topology-preserving mask injection with edge gating to maintain thin-structure continuity and suppress false positives. A class-conditional flow-matching mask model synthesizes masks with control over crack coverage, enabling balanced, topology-diverse data without manual annotation. We inject masks into crack-free backgrounds to diversify illumination and reduce false positives. On five datasets with a CNN-Transformer backbone, incorporating synthesized pairs improves in-domain performance by 5.37 mIoU and 5.13 F1, and target-guided cross-domain synthesis yields gains of 13.12 mIoU and 14.82 F1 using target mask statistics. We also release CSF-50K, 50,000 image-mask pairs for benchmarking.

Figures (18)

• Figure 1: Given mask $y$, sample $x_0\!\sim\!\mathcal{N}(0,I)$, $x_1\!\sim\!p_{\text{data}}(\cdot\mid y)$, $\xi\!\sim\!\mathcal{N}(0,I)$. Interpolant $x_t=\phi_t(x_0,x_1,\xi)=(1-t)x_0+tx_1+g(t)\xi$, $g(0)=g(1)=0$. Oracle velocity $v^*(x_t,t)=\partial_t\phi_t(x_0,x_1,\xi)$; train $v_\theta(x_t,t,y)$ to regress $v^*$. Sampling solves the ODE $\frac{\mathrm d x_t}{\mathrm dt}=v_\theta(x_t,t,y)$, i.e., $x_1=x_0+\int_0^1 v_\theta(x_t,t,y)\,\mathrm dt$.
• Figure 2: Noise to Image with Crack.
• Figure 3: U -Net backbone used in our branch.
• Figure 4: CrackSegFlow residual block used in the velocity-field network $v_\theta$.
• Figure 5: Topology-Preserving Mask Injection paired with Boundary-Gated Modulation. The former is our modified SPADE-style conditioning applied at every decoder block to preserve mask topology across scales, while the latter selectively amplifies features near crack boundaries to recover sub-pixel filaments.