Morphology-Decoupled Segmentation for Multiple Tunnel Defects Under Complex Environments
Authors: Mohan Ren, Hang Li, Yifei Liu, Shaowei Wang, James O'Kane, Xuanzhi Wang, Shiyu Zhao, Yumeng Zhang, Songtao Ji, Jurij Karlovsek
Organizations: School of Civil Engineering, The University of Queensland, St. Lucia, Brisbane, 4072, Queensland, Australia · Department of Computer Science, University of Exeter, The Queen’s Drive, Exeter, EX4 4QJ, Devon, United Kingdom · School of Civil and Environmental Engineering, Nanyang Technological University, , Singapore, 639798, Singapore · School of Civil Engineering, Chang’an University, , Xi’an, 710064, Shaanxi, China · Xi’an University of Science and Technology, , Xi’an, Shaanxi, China
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.
Figures & tables
Figure 1: Dataset composition and multi-defect examples. (a) Of 1,100 images and 3,089 instances, 99 images (9.0%) contain multiple classes. (b) Crack, leakage, and lining fall-off. (c) Crack, leakage, and segment damage. (d) Leakage B/W with segment damage. (e) Crack crossing areal leakage. (f) Crack with Leakage G. (g) Leakage G with small fall-off regions. Colours follow the legend; arrowheads mark selected defects. Images provided by the organiser.
Class
Instances
Instance Share
Pixel Share
Leakage B
1,412
45.71%
72.21%
Crack
759
24.58%
0.36%
Lining falling off
556
18.00%
4.73%
Segment damage
213
6.90%
16.59%
Leakage W
91
2.95%
1.88%
Leakage G
58
1.88%
4.23%
Table 1: Class distribution of tunnel lining defects.
Figure 2: Descriptor distributions by class: (a) compactness, (b) solidity, and (c) elongation. Violin width shows density, black bars mark medians, and red frames highlight cracks, which separate from the five areal categories.
Crack Width
IoU 0 px
IoU 1 px
IoU 2 px
Interpretation
2 px
1.000
0.333
0.000
Almost no tolerance
3 px
1.000
0.500
0.200
1 px error halves IoU
5 px
1.000
0.667
0.429
Still highly sensitive
10 px
1.000
0.818
0.667
Moderate tolerance only
Table 2: Crack IoU sensitivity to lateral spatial shifts, computed from Eq. ( 1 ).
Group
Pixel Range
Processing
Stride
Tiny
< 200 K
Resize to 5122
—
Mobile
200 K–2 M
5122 tiles
384
DSLR
2 M–8 M
6402 tiles → input
400
HighRes
> 8 M
7682 tiles → input
384
Table 3: Resolution groups and their processing strategies.
Figure 3: Resolution-adaptive inference. (a) Workflow from resolution grouping and overlapping tiles to 512×512 inference, probability merging, and reconstruction; HighRes uses 768×768 tiles with stride 384. (b) A 12.0-MP image. (c) Its 70 tiles at 50% overlap. (d) Gaussian-weighted full-resolution Leakage G prediction.
Figure 4: MDS architecture. (a) DINOv3-distilled ConvNeXt-Tiny backbone and features C1–C4. (b) MRM routing into complementary linear and areal streams. (c) DSA and PPM-FPN decoders producing FL and FA . (d) CMIM cross-attention, gated refinement, main prediction, and training-only auxiliary heads.
Figure 5: Supervised routing for (a,c) crack with Leakage B and (b) crack with lining fall-off. Rows show input, ground truth, morphology target, routing map α , and MDS prediction. Red/blue denote linear/areal targets; grey background is ignored. Low/high α favours the areal/linear stream. μL , μA , Δ , and foreground IoU are reported below.
DSA Decoder (Linear)
Areal Decoder
Target morphology
Cracks ( >10× elong.)
Leakage, spalling, damage
Core mechanism
Strip attention
Pyramid Pooling
Receptive field
Directional, adaptive
Isotropic, multi-scale
Output
[B,32,4H,4W]
[B,32,4H,4W]
Table 4: Deliberate design asymmetry between the two decoders, reflecting the morphological dichotomy established in Section 2 .
Hyper-parameter
Value
Backbone
DINOv3 ConvNeXt-Tiny
Network input
512×512
Decoder channel
32
Optimiser
AdamW
Weight decay
1e-2
Batch size
32
Table 5: Hyper-parameter settings for MDS.
Model
mIoU (%)
mDice (%)
aAcc (%)
DeepLabv3+
54.27
67.48
84.66
UNet
23.35
31.82
76.72
SegFormer
76.20
86.43
95.87
PSPNet
71.48
81.67
94.63
SegDINO
80.08
87.87
97.87
MDS (ours)
81.17
88.47
98.11
Table 6: Overall quantitative comparison under the same validation and evaluation protocol; encoder initialisation follows the model-specific settings in Section 4.1 .
Class
Baseline
MDS- Full
w/o CMIM
w/o Routing
Background
97.62
97.88
97.71
97.71
Crack
44.67
43.86
47.18
46.08
Leakage B
81.57
83.97
82.74
82.51
Leakage W
72.39
72.63
71.81
72.24
Leakage G
91.41
92.45
90.24
91.90
Lining Falling Off
80.11
83.59
80.19
79.72
Table 7: Per-class IoU comparison of the backbone-matched baseline, full MDS, and component ablations.
Class
SegDINO F1 (%)
MDS F1 (%)
Crack
88.57
87.67
Leakage B
97.32
96.71
Leakage W
53.33
71.43
Leakage G
50.00
50.00
Lining Falling Off
74.07
83.33
Segment Damage
87.50
93.02
Table 8: Per-class image-level F1 comparison for foreground defect detection.
Figure 6: Nine validation cases: (a) input, (b) ground truth, and predictions from (c) U-Net, (d) PSPNet, (e) DeepLabV3+, (f) SegFormer, (g) SegDINO, and (h) MDS. Values are per-image mIoU; the red frame marks MDS.
Metric (%)
Baseline
MDS- Full
w/o CMIM
w/o Routing
mIoU
80.08
81.17
80.47
80.49
mIoU fg
77.16
78.39
77.60
77.62
mDice
87.87
88.47
88.23
88.18
mDice fg
86.05
86.73
86.46
86.40
mAcc
86.16
86.64
86.29
85.56
mPrecision
89.94
90.79
90.55
91.30
Table 9: Aggregate ablation results for the MDS components under the same evaluation protocol.
Metric (%)
Baseline
MDS- Full
w/o CMIM
w/o Routing
Crack IoU
44.67
43.86
47.18
46.08
Crack Dice
61.75
60.98
64.12
63.09
Crack recall
59.95
57.27
61.41
58.05
Crack precision
63.66
65.21
67.07
69.08
Crack image F1
88.57
87.67
91.67
91.18
Crack component F1
78.74
74.61
78.48
79.88
Table 10: Crack-specific pixel-, image-, and topology-sensitive diagnostic metrics.
Configuration
Leakage mIoU (%)
Leakage mDice (%)
G → B Conf.
Baseline
81.79
89.78
0.013%
MDS-Full
83.02
90.50
0.000%
w/o CMIM
81.60
89.67
0.149%
w/o Routing
82.21
90.03
0.745%
Table 11: Leakage-subclass metrics and Leakage G → Leakage B confusion under different configurations.
Module
Parameters
DINOv3 ConvNeXt-Tiny backbone
27,818,592
MRM
1,634,689
Linear/DSA decoder
88,768
Areal decoder
443,136
CMIM
12,736
Main classification head
455
Table 12: Exact trainable-parameter breakdown of the full MDS model.
Configuration
FP32 size (MiB)
Dynamic INT8 size (MiB)
Compression ratio
FP32 mIoU
Dynamic INT8 mIoU
Δ mIoU
Baseline
106.4
32.4
3.29×
0.8008
0.7984
−0.0024
MDS (full)
114.5
40.5
2.83×
0.8117
0.8108
−0.0009
Table 13: Checkpoint size and segmentation accuracy before and after selective dynamic INT8 quantisation.
Figure 7: Failure cases: (a) input, (b) ground truth, and predictions from (c) U-Net, (d) PSPNet, (e) DeepLabV3+, (f) SegFormer, (g) SegDINO, and (h) MDS. Values are per-image mIoU; the red frame marks MDS. L219, L163, and S41 illustrate faint Leakage W, Leakage B/W confusion, and a crack beside segment damage.
Tunnel inspection requires outputs that can support defect localization, measurement, severity grading, and engineering documentation. Existing training-free foundation-model pipelines usually stop at coarse open-vocabulary proposals, which are difficult to use directly in interference-heavy tunnel scenes. We propose a training-free framework TunnelMIND. Specifically, language-guided defect proposals are not treated as final outputs; instead, their spatial support is recalibrated at inference time through dense visual consistency, so that coarse semantic anchors can be transformed into more reliable prompts under tunnel-specific hard negatives. The resulting masks are further reconstructed into structured defect entities with category, location, geometry, severity, and context attributes, which are then mapped to retrieval-grounded explanation and engineering-readable report generation under expert knowledge constraints. On visible, GPR, and road defect tasks, TunnelMIND achieves F1 scores of 0.68, 0.78, and 0.72, respectively. Overall, TunnelMIND shows that training-free tunnel inspection can move beyond coarse localization toward structured defect evidence for engineering assessment.
Shipeng Liu, Liang Zhao, Dengfeng Chen +1
Xi’an University of Architecture and Technology, School of mechanical and electrical engineering, Xi’an, 710055, Shaanxi, China · Xi’an University of Architecture and Technology, School of Information and Control Engineering, Xi’an, 710055, Shaanxi, China · Xi’an University of Architecture and Technology, School of Civil Engineering, Xi’an, 710055, Shaanxi, China
Recent crack segmentation methods often follow generic semantic segmentation designs, using stronger backbones, hybrid CNN-Transformer-Mamba encoders, and auxiliary enhancement branches. Although effective, this raises whether stronger generic feature mixing is the most suitable direction for crack segmentation. We instead formulate crack segmentation as sparse structural recovery. Cracks have limited category-level semantics but strong morphological regularities, being thin, sparse, anisotropic, locally fragmented, and easily confused with textures or shadows. Thus, the key bottleneck lies in preserving weak structural evidence, recovering directional continuity, and suppressing background coupling. We propose RIFT, a compact family of morphology-aligned crack segmentation models. Rather than compressing a complex generic architecture, RIFT is simple by design, preserving local evidence, aggregating cooperative directional continuity, and restoring crack structures through lightweight multi-scale fusion. Experiments on four public benchmarks show that RIFT achieves the best or tied-best results across the 16 main metrics against reproduced representative baselines. RIFT-B gives the strongest overall accuracy, while RIFT-T provides the best deployment efficiency with only 0.47M parameters and high inference speed. Topology-aware evaluation, ablations, transfer experiments, and visualizations further verify that task-aligned simplicity can match or surpass complex hybrid architectures when its inductive bias fits crack morphology. Code: https://github.com/xauat-liushipeng/RIFT
Feature encoders play a key role in pixel-level crack segmentation by shaping the representation of fine textures and thin structures. Existing CNN-, Transformer-, and Mamba-based models each capture only part of the required spatial or structural information, leaving clear gaps in modeling complex crack patterns. To address this, we present MixerCSeg, a mixer architecture designed like a coordinated team of specialists, where CNN-like pathways focus on local textures, Transformer-style paths capture global dependencies, and Mamba-inspired flows model sequential context within a single encoder. At the core of MixerCSeg is the TransMixer, which explores Mamba's latent attention behavior while establishing dedicated pathways that naturally express both locality and global awareness. To further enhance structural fidelity, we introduce a spatial block processing strategy and a Direction-guided Edge Gated Convolution (DEGConv) that strengthens edge sensitivity under irregular crack geometries with minimal computational overhead. A Spatial Refinement Multi-Level Fusion (SRF) module is then employed to refine multi-scale details without increasing complexity. Extensive experiments on multiple crack segmentation benchmarks show that MixerCSeg achieves state-of-the-art performance with only 2.05 GFLOPs and 2.54 M parameters, demonstrating both efficiency and strong representational capability. The code is available at https://github.com/spiderforest/MixerCSeg.
Zilong Zhao, Zhengming Ding, Pei Niu +2
School of Qilu Transportation, Shandong University, China · Department of Computer Science, Tulane University, USA