Morphology-Decoupled Segmentation for Multiple Tunnel Defects Under Complex Environments

Mohan Ren, Hang Li, Yifei Liu, Shaowei Wang, James O'Kane, Xuanzhi Wang, Shiyu Zhao, Yumeng Zhang, Songtao Ji, Jurij KarlovsekarXiv 2026 · arXiv

TL;DR

Researchers have developed Morphology-Decoupled Segmentation (MDS), a deep learning framework that treats thin, linear cracks and spatially extensive "areal" defects (like leakage or spalling) as fundamentally different geometric problems. By routing features into specialized decoders and using gated cross-morphology attention, the model achieved second place in a major tunnel inspection competition, balancing high-precision crack connectivity with broad regional context.

What Problem Does This Paper Address?

Automated tunnel inspection is a nightmare for standard computer vision models. Tunnels suffer from uneven lighting, background clutter (staining, shadows), and a massive "morphological dichotomy" among defects.

On one hand, you have cracks: thin, elongated structures where a single pixel shift in prediction can halve the Intersection over Union (IoU) score. On the other, you have areal defects like water leakage or concrete spalling: diffuse, region-based blobs that require broad spatial context to identify.

Existing models usually apply a "one-size-fits-all" decoder. If the model is biased toward cracks, it might miss the diffuse boundaries of a leak; if it's biased toward regions, it often "fragments" thin cracks. The authors tackle this by making morphology—the actual shape of the defect—the organizing principle of the network architecture.

Methods

The MDS framework introduces several novel components to handle this geometric split:

  1. Morphological Routing Module (MRM): Instead of a standard feature pyramid, the MRM uses a supervised soft-gating mechanism to partition encoder features into two streams: Linear and Areal.
  2. Linear Decoder (Deformable Strip Attention): Cracks don't follow a grid. This module predicts direction hypotheses for cracks and uses "strip attention" to aggregate evidence along oriented paths rather than fixed horizontal or vertical axes. This preserves crack continuity.
  3. Areal Decoder (PPM-FPN): This uses a Pyramid Pooling Module to capture the multi-scale context needed for large, diffuse defects like water stains.
  4. Cross-Morphology Interaction Module (CMIM): The two streams aren't isolated. CMIM uses gated cross-attention to allow the branches to exchange information (e.g., a crack might be the source of a leak) without losing their specialized representations.
  5. Resolution-Adaptive Pipeline: To prevent thin cracks from disappearing during resizing, the authors implemented a tiling strategy that adjusts based on the native resolution of the input image (from "Tiny" to "HighRes").
  6. Model Compression: For real-world deployment, they applied selective dynamic INT8 quantization, targeting the linear layers of the backbone and decoders.

Datasets

The study utilizes a dataset of 1,100 tunnel-lining images containing 3,089 annotated defect instances.

  • Classes: Six foreground categories—Crack, Leakage B (black stain), Leakage W (white efflorescence), Leakage G (algae), Lining falling off, and Segment damage.
  • Challenges: The dataset is highly imbalanced; cracks represent 24.5% of instances but only 0.36% of the total pixels, highlighting the extreme difficulty of the segmentation task.

Key Insights and Ideas

  • Supervised Routing is Essential: The authors found that simply splitting the network isn't enough; you have to tell the network which features are linear and which are areal during training. This prevents the branches from collapsing into redundant copies of each other.
  • The Crack-Areal Trade-off: Ablation studies revealed a fascinating tension. Removing the interaction module (CMIM) actually improved crack IoU but hurt every other category. This suggests that while cracks benefit from extreme specialization, the overall health of a multi-defect model requires "controlled interference" between streams.
  • Directional Hypotheses: By allowing the model to "guess" the orientation of a crack and sample points along that line, the model acts more like a human engineer tracing a crack by eye, rather than a standard filter looking at a static 3x3 box.

Key Results

  • State-of-the-Art Performance: MDS achieved 81.17% mIoU on the internal validation set, outperforming the backbone-matched SegDINO baseline by 1.09 points.
  • Competition Success: In the 2026 Intelligent Multi-defect Detection for Tunnel Linings competition, the model reached 67.55% mIoU on a hidden test set, securing second place.
  • Effective Compression: The INT8 quantization reduced the model size from 114.5 MiB to 40.5 MiB (a 2.83x reduction) with a negligible mIoU loss of only 0.0009.
  • Image-Level Recognition: The model significantly improved the "foreground image-level F1 score" (a measure of whether the model correctly identifies if a defect is present at all), which is critical for automated screening pipelines.

Limitations and Future Work

Despite the gains, the authors noted several remaining hurdles:

  • Chromatic Confusion: The model still struggles to distinguish between different types of leakage (e.g., black vs. white) when lighting is poor.
  • Faint Defects: Very faint water stains (low contrast) are still frequently missed by all tested methods.
  • Small Cracks Near Large Damage: When a thin crack exists right next to a massive area of segment damage, the "louder" signal of the large defect can sometimes suppress the crack detection.
  • Future Directions: The authors suggest moving toward a three-regime routing system (adding a "point-like" or "intermediate" category) and incorporating explicit color-aware modeling to better handle leakage subclasses.

For more details, see the full paper: "Morphology-Decoupled Segmentation for Multiple Tunnel Defects Under Complex Environments" (arXiv:2611.19981v1).

Summary updated

Abstract

Automated tunnel-lining inspection must segment different defect geometries: cracks are thin, elongated, and connectivity-sensitive, whereas leakage, lining falling off, and segment damage are spatially extensive and context-dependent. Existing approaches generally apply a common decoder across defect classes or introduce boundary, topology, and directional mechanisms as isolated enhancements. This paper introduces Morphology-Decoupled Segmentation (MDS), in which defect morphology becomes the organising principle for feature routing, decoding, and interaction. A supervised routing module partitions multiscale encoder features into complementary linear and areal streams. The linear stream employs Deformable Strip Attention with image-conditioned direction hypotheses to preserve crack continuity without assuming fixed horizontal or vertical orientations, while the areal stream combines Pyramid Pooling with feature-pyramid fusion to capture diffuse boundaries and regional context. The representations are coupled through gated cross-morphology attention, allowing complementary information exchange while retaining morphology-specific features. A resolution-adaptive tiling and probability-blending pipeline preserves fine defects across variable-resolution imagery. MDS achieves 81.17% mIoU on the internal validation set and improves foreground image-level recognition over a backbone-matched SegDINO model, primarily for areal defects. Controlled ablations demonstrate the complementary contributions of routing supervision and cross-stream interaction while revealing a trade-off between areal-class performance and crack specialisation. The submitted model achieved 67.55% mIoU on an organiser-controlled hidden test, and the team placed second. Selective dynamic INT8 quantisation reduced model storage by 2.83 times with negligible accuracy loss.
BibTeX
@misc{ren2026morphology,
  author = {Mohan Ren and Hang Li and Yifei Liu and Shaowei Wang and James O'Kane and Xuanzhi Wang and Shiyu Zhao and Yumeng Zhang and Songtao Ji and Jurij Karlovsek},
  title = {{Morphology-Decoupled Segmentation for Multiple Tunnel Defects Under Complex Environments}},
  year = {2026},
  eprint = {2611.19981},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19981},
}
Figure