DROA-CLIPSeg: Prompt-Guided Thin Crack Segmentation in Low-Light Conditions

Vinita Shukla; Partibha Jangra; Vinayak Pathak; Praful Hambarde; Amit Shukla

cs.CVOct 24, 2026

DROA-CLIPSeg: Prompt-Guided Thin Crack Segmentation in Low-Light Conditions

Authors: Vinita Shukla, Partibha Jangra, Vinayak Pathak, Praful Hambarde, Amit Shukla

Organizations: Indian Institute of Technology Mandi · Indian Institute of Technology Kanpur

Abstract

Pavement crack segmentation under low light is degraded by reduced contrast, shadows, headlight glare, joints, stains, and wet seams. Existing crack segmentation methods mostly assume good illumination. Low-light enhancement methods improve visual quality, but they do not explicitly preserve crack topology. To address these limitations, we formulate low-light crack segmentation as a degradation-aware, prompt-grounded structural defect segmentation problem and propose DROA-CLIPSeg, a parameter-efficient framework that adapts a frozen CLIPSeg vision language backbone using two lightweight modules. QKV Low-Rank Adaptation is injected into the Q/K/V projections of the vision and text encoders to improve crack-specific image text alignment without backbone fine-tuning. The proposed Domain Residual Orthogonal Adapter (DROA) decomposes intermediate features into a dominant crack-domain branch, which captures recurring low-light crack patterns, and an orthogonal residual branch, which preserves faint, thin, and fragmented evidence usually suppressed by noise, shadow, or over-smoothing. We also introduce CrackLingo-9K and CrackLingo-12K, text augmented versions of Crack9k and Crack12k with crack-aware prompts. Training only approximately 1.93 M parameters, corresponding to 1.3% of the full model, DROA-CLIPSeg outperforms 9 zero-shot and 4 fine-tuned vision language baselines on both splits, generalises zero-shot to MVTec AD across textures and defect morphologies, and is supported by extensive ablations. Our source code is publicly available at https://github.com/VShuklacode/DROA-CLIPSeg

Figures & tables

Figure 1: Overall architecture of DROA-CLIPSeg: frozen CLIP encoders adapted by QKV-LoRA and DROA modules, fused by the CLIPSeg decoder and a refinement head.

Figure 2: QKV Low-Rank Adaptation: a trainable low-rank pair $(B,A)$ injects $\Delta W\!=\!BA$ into the frozen $Q,K,V$ projections.

Panel of Figure 3 — Figure 3: (a) DROA: dominant + orthogonal-residual branches fused by a learnable energy gate. (b) CrackLingo-12K-trained DROA-CLIPSeg applied zero-shot to MVTec-AD (CVPR’19) [ 2 ] Grid Thread / Wood / Tile / Leather (columns: Input, GT, Heatmap, Overlay, Pred Mask).

CrackLingo-9K ( $N\!=\!395$ )
Model	Img-A	Pix-A	Dice(px)	Dice(img)
CLIPSeg (CVPR’22) [ 28 ]	52.76	52.76	7.75	4.46
LSeg (ICLR’22) [ 19 ]	53.11	53.11	5.46	3.69
DenseCLIP † (CVPR’22) [ 34 ]	11.34	11.34	6.39	6.06
MaskCLIP (ECCV’22) [ 47 ]	28.74	28.74	7.18	7.23
OV-Seg (CVPR’23) [ 23 ]	38.78	38.78	6.58	6.67

Table 1: Tier-1 zero-shot image-level results on CrackLingo-9K/-12K ; %. † CLIP-proxy

Figure 4: CrackLingo prompt format: per-sample JSON with prompt composed over orientation, vision, location, category axes.

CrackLingo-9K ( $N\!=\!395$ )
Model	F1	IoU	Prec	Rec	Acc
CLIPSeg (CVPR’22) [ 28 ]	7.75	4.03	4.12	64.17	52.76
LSeg (ICLR’22) [ 19 ]	5.46	2.81	2.91	43.85	53.11
DenseCLIP † (CVPR’22) [ 34 ]	6.39	3.30	3.30	97.91	11.34
MaskCLIP (ECCV’22) [ 47 ]	7.18	3.72	3.74	89.17	28.74
OV-Seg (CVPR’23) [ 23 ]	6.58	3.40	3.45	69.77	38.78

Table 2: Tier-1 zero-shot results on CrackLingo-9K/-12K ; %. † CLIP-proxy

Split	Model	Train	Pixel metrics					Anomaly-detection metrics
Split	Model	Train	F1	IoU	Prec	Rec	Acc	p-AUC	p-F1	p-AP	i-AUC	i-F1	i-AP
9K	CLIPSeg-FT (CVPR’22) [ 28 ]	dec	47.97	31.55	38.35	64.02	95.70	90.17	48.71	45.26	75.86	91.46	93.37
	MaskCLIP-FT (ECCV’22) [ 47 ]	dec+L2	50.58	33.85	39.08	71.68	95.67	92.97	56.27	56.26	92.93	93.98	98.60
	SAN-headFT ∗ (CVPR’23) [ 42 ]	head	4.63	2.37	3.44	7.06	91.00	59.85	7.75	3.81	76.63	91.52	94.51
	CAT-Seg-headFT ∗ (CVPR’24) [ 8 ]	head	7.74	4.03	6.16	10.41	92.33	62.80	9.19	5.10	66.46	91.33	92.51
	DROA-CLIPSeg (ours)	L+D	70.75	54.74	65.30	77.20	96.76	95.13	66.87	63.97	98.97	97.65	99.81

Table 3: Tier-2 fine-tune results on CrackLingo-9K (

N\!=\!395

) and -12K (

N\!=\!1200

)

Method	Params	CrackLingo-12K					CrackLingo-9K
Method	Params	F1	IoU	Prec	Rec	Acc	F1	IoU	Prec	Rec	Acc
CLIPSeg zero-shot (CVPR’22) [ 28 ]	150.7 M	16.47	16.44	94.74	18.49	91.98	7.88	7.75	90.63	8.50	91.57
CLIPSeg end-to-end FT (CVPR’22) [ 28 ]	149.8 M	69.19	58.78	69.35	73.96	96.15	51.02	38.68	55.39	60.66	96.43
LoRA fine-tune (ICLR’22) [ 13 , 28 ]	1.1 M	62.82	52.79	62.42	70.18	95.58	47.68	36.54	55.40	55.10	95.39
AA-CLIP (CVPR’25) [ 29 ]	50.3 M	72.25	56.56	69.23	76.00	96.05	60.16	43.02	58.16	62.31	92.48
ResNet50 1-shot (CVPR’16) [ 11 ]	64.0 M	41.94	43.11	50.08	50.21	88.40	49.94	45.67	49.94	49.94	88.27

Table 4: Full method comparison on CrackLingo-12K and -9K.

Panel of Figure 5 — Figure 5: (a) Qualitative crack-segmentation comparison (zero-shot): (a) Input Image, (b) GT, (c) zero-shot CLIPSeg [ 28 ] , (d) LSeg [ 19 ] , (e) DenseCLIP [ 34 ] , (f) MaskCLIP [ 47 ] , (g) OV-Seg [ 23 ] , (h) SAN [ 42 ] , (i) ZegCLIP [ 48 ] , (j) CAT-(k) Seg [ 8 ] and their fine-tuned counterparts (l) CLIPSeg-FT [ 28 ] , (m) MaskCLIP-FT [ 47 ] , (n) CAT-Seg-head-FT [ 8 ] , DROA-CLIPSeg(ours). (b) Fine-tuned baseline comparison on CrackLingo-9K/-12K. (a) input, (b) zero-shot CLIPSeg [ 28 ] , (c) end-to-end CLIPSeg-FT, (d) LoRA-CLIPSeg-FT, (e) AA-CLIP [ 29 ] , (f–i) ResNet50/101 1-/5-shot [ 11 ] , (j) GT, (k) DROA-CLIPSeg (ours).

Figure 6: Text-contribution analysis on CrackLingo-12K under seven prompt perturbations ( our text , coarse crack prompts , empty text , wrong no-crack , wrong thin , wrong thick , shuffled text ).

#	Ablation	Modification	DSC	IoU	P	R	thin	thick	no-crk
A. LoRA rank [ 13 ]
1	lora_r4	LoRA $r=4$	69.50	59.20	67.73	81.28	47.25	73.43	93.40
2	lora_r8	LoRA $r=8$	69.27	58.78	65.29	82.71	48.87	73.24	90.09
3	lora_r16	LoRA $r=16$	68.85	58.45	66.78	80.34	48.21	72.95	89.62
4	lora_r32	LoRA $r=32$	69.79	59.47	67.82	81.27	48.52	72.69	95.28
5	lora_r64	LoRA $r=64$	69.74	59.35	67.39	82.11	49.24	73.60	91.04

Table 6: Ablation summary on CrackLingo-12K test split (%,

n{=}1200

Crack Segmentation

Crack segmentation locates cracks at the pixel level in images to support structural inspection, damage measurement, and materials analysis.

See all 19 papers

Rethinking Efficient Crack Segmentation with Task-Aligned Structural-Directional Modeling
May 29, 2026Shipeng Liu, Liang Zhao, Dengfeng Chen +1Efficient Semantic Segmentation Crack Segmentation
Xi’an University of Architecture and Technology
Semantic-Edge Response Decoding of SAM3 for Zero-Shot Crack Segmentation
Jul 14, 2026Shipeng Liu, Zhanping Song, Liang Zhao +1Promptable Image Segmentation Boundary-Aware Segmentation
Xi’an University of Architecture and Technology
ELVIS: Enhance Low-Light for Video Instance Segmentation in the Dark
Dec 1, 2025Joanne Lin, Ruirui Lin, Yini Li +2Video Instance Segmentation Instance Segmentation
Visual Information Laboratory, University of Bristol

May 29, 2026J/K move · Enter open · S save

Rethinking Efficient Crack Segmentation with Task-Aligned Structural-Directional Modeling

Shipeng Liu, Liang Zhao, Dengfeng Chen, Weihua Zhang

Xi’an University of Architecture and Technology

Efficient Semantic Segmentation Crack Segmentation Semantic Segmentation Thin-Structure Segmentation

Open paper AskPDF arXiv

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

Abstract

Crack Segmentation

Explore similar work

Rethinking Efficient Crack Segmentation with Task-Aligned Structural-Directional Modeling

Semantic-Edge Response Decoding of SAM3 for Zero-Shot Crack Segmentation

ELVIS: Enhance Low-Light for Video Instance Segmentation in the Dark