Research map / Detection and segmentation / CNN features
Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CIRL, a two-phase framework that learns per-pixel segmentation variance (Aseg Loss, Wasserstein-based) to localize ambiguous marginal non-crack regions, then applies a clustering-inspired loss (CI Loss) to those regions for crack segmentation.
Problem
Crack segmentation is hurt by ambiguous labels in marginal non-crack regions, which have low contrast and heterogeneous texture, making discriminative feature learning hard.
Approach
Learn a predicted segmentation variance via a Wasserstein-distance loss to identify ambiguous marginal non-crack pixels, then replace supervised BCE on those pixels with an unsupervised neighbor-consistency clustering loss that encourages nearby features to make consistent predictions and distant features to be inconsistent.
Key results
- JTFN + CIRL reaches 86.53 F1 on CrackTree200 (vs 84.19 for JTFN) and 67.62 F1 on Crack500 (vs 65.76).
- On CrackSeg5k, JTFN + CIRL achieves 60.08 F1 vs 58.42 for JTFN and 58.83 for Crackformer.
- CI Loss beats NNCLR on CrackSeg5k with JTFN (60.08 vs 58.95 F1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CIRL to JTFN | JTFN · CrackTree200 | F1: 84.19 -> 86.53 | Helped |
| Add CIRL to JTFN | JTFN · Crack500 | F1: 65.76 -> 67.62 | Helped |
| Add CIRL to JTFN | JTFN · DRIVE | F1: 82.81 -> 84.32 | Helped |
| Add CIRL to JTFN | JTFN · CrackSeg5k | F1: 58.42 -> 60.08 | Helped |
| Add Phase II, including CI Loss, after Phase I (ablation) | JTFN with Phase I · CrackSeg5k | F1: 58.57 -> 60.08 | Helped |
| Replace NNCLR with CI Loss (ablation) | JTFN · CrackSeg5k | F1: 58.95 -> 60.08 | Helped |
Surprising findings
- Excluding marginal non-crack regions from BCE training changes F1 inconsistently across sub-training sets (-3.12% to +3.48%), showing direct use of these labels can be harmful.
Limitations
- Two-phase schedule with a fixed 1000-epoch switch and several hyperparameters (lambda, gamma, K, beta) to tune.
- Evaluated mainly with JTFN as base model; generality to other architectures is only partly shown.
Experimental setup
Datasets: CrackSeg5k Crack500 CrackTree200 DRIVE
Benchmarks: CrackSeg5k Crack500 CrackTree200 DRIVE
Metrics: F1 Precision Recall
Baselines: UNet VGG-UNet TopoNet DRU Crackformer JTFN Focal Loss Robust Dice Loss clDice Loss Poly Loss Robust T-Loss NNCLR
Builds on: JTFN Crackformer NNCLR Wasserstein distance
Closest papers by meaning
- SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures (CVPR 2025 · similarity 0.85)
- Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models (NeurIPS 2026 · similarity 0.81)
- SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.80)
- Attack-inspired Calibration Loss for Calibrating Crack Recognition (AAAI 2025 · similarity 0.79)
- Scribbles for All: Benchmarking Scribble Supervised Segmentation Across Datasets (NeurIPS 2024 · similarity 0.79)
- Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.