MixerCSeg: An Efficient Mixer Architecture for Crack Segmentation via Decoupled Mamba Attention

Zilong Zhao; Zhengming Ding; Pei Niu; Wenhao Sun; Feng Guo

cs.CVMar 2, 2026

MixerCSeg: An Efficient Mixer Architecture for Crack Segmentation via Decoupled Mamba Attention

Authors: Zilong Zhao, Zhengming Ding, Pei Niu, Wenhao Sun, Feng Guo

Organizations: School of Qilu Transportation, Shandong University, China · Department of Computer Science, Tulane University, USA

Abstract

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.

Figures & tables

Figure 1: Visualization results of heatmaps and attention maps in VMamba [ 18 ] . Benefiting from the multi-directional scanning mechanism, tokens in global channels can achieve a receptive field that covers the entire scope, rather than being limited to historical context. In contrast, local channels focus on the field of view of neighboring regions.

Figure 2: Overview of the proposed method. (a) illustrates the overall architecture of MixerCSeg. TransMixer blocks and downsampling operators are connected in series to extract multi-scale features. Before decoding, the DEGConv module is employed to enhance crack texture details. (b) demonstrates the design of the DEGConv, which integrates a spatial block strategy and directional prior knowledge.

Figure 3: The difference between TransMixer and existing methods. (a) illustrates two common patterns in existing methods that hybridize Mamba blocks and Transformer blocks. (b) shows the design details of the TransMixer module, which decomposes features into global and local representations, enhancing them with Self-Attention and Local Refinement, respectively.

Method	DeepCrack				CamCrack789				CrackMap				Crack500
Method	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
U-Net	0.8987	0.8868	0.8940	0.8999	0.8372	0.7991	0.8066	0.7353	0.7983	0.7540	0.7865	0.7363	0.7105	0.6238	0.6610	0.5410
CarNet	0.8875	0.8688	0.8702	0.8939	0.8257	0.7830	0.7853	0.7251	0.7568	0.6840	0.7117	0.6400	0.6428	0.5487	0.5921	0.3636
RINDNet	0.8391	0.8087	0.8267	0.8377	0.8137	0.7628	0.7728	0.6050	0.7425	0.6745	0.6943	0.6699	0.7381	0.6469	0.6483	0.7119
DTrCNet	0.8661	0.8473	0.8512	0.8566	0.8150	0.7719	0.7749	0.7836	0.7812	0.7328	0.7413	0.7276	0.7627	0.7012	0.7241	0.7357
SCSegamba	0.9022	0.8938	0.8990	0.9110	0.8268	0.7919	0.7939	0.7876	0.8094	0.7741	0.7766	0.7678	0.7778	0.7244	0.7370	0.7553

Table 1: Performance comparison on four crack datasets. The best result for each metric is in bold.

Method	FLOPs (G)	Params (M)	Memory (MiB)
U-Net	204.38	28.99	4394
CarNet	19.00	4.89	1824
RINDNet	695.77	59.39	5392
DTrCNet	122.55	63.83	4952
SCSegamba	18.16	2.80	2206
RestorMixer	98.71	3.19	10384

Table 2: Comparison of computational cost.

Figure 4: Visualization results of MixerCSeg versus state-of-the-art methods in various environments.

TransMixer	DEGConv	SRF	FLOPs (G)	Params (M)	Memory (MiB)	DeepCrack				CamCrack789
TransMixer	DEGConv	SRF	FLOPs (G)	Params (M)	Memory (MiB)	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
			17.74	0.76	1986	0.8826	0.8696	0.8753	0.8867	0.8283	0.7928	0.8007	0.8051
✓			19.10	2.50	3594	0.9016	0.8931	0.9045	0.9092	0.8359	0.8041	0.8164	0.8163
✓	✓		19.18	2.64	3626	0.9097	0.9032	0.9127	0.9107	0.8381	0.8073	0.8147	0.8211
✓	✓	✓	2.05	2.54	1190	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244

Table 3: Component ablation experiments on the DeepCrack and CamCrack789 datasets.

Encoder	DeepCrack				CamCrack789
Encoder	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
MambaVision	0.9020	0.8936	0.8979	0.9081	0.8307	0.7964	0.8132	0.8131
RestorMixer	0.9087	0.9022	0.9122	0.9142	0.8373	0.8063	0.8206	0.8215
TransMixer (Ours)	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244

Table 4: Ablation experiments of different hybrid blocks on the DeepCrack and CamCrack789 datasets.

Value	DeepCrack				CamCrack789
Value	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
$\gamma=0.5$ , cell size is $(8,8)$ and $n=180$
-	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244
The ratio $\gamma$ of global channels
0.3	0.9098	0.9035	0.9117	0.9172	0.8375	0.8065	0.8240	0.8243
0.7	0.9133	0.9073	0.9182	0.9114	0.8379	0.8072	0.8171	0.8231

Table 5: Analysis of Hyperparameter Experiments of the MixerCSeg Model on the DeepCrack and CamCrack789 Datasets.

Method	mIoU	ODS	OIS	F1
ARConv [ 21 ]	0.9106	0.9043	0.9093	0.9170
Ours	0.9151	0.9094	0.9197	0.9205

Table 6: Results of direction embeddings generated based on ARConv-predicted angles and those generated by the proposed method on the DeepCrack dataset.

Local	Global	DeepCrack				CamCrack789
Local	Global	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
		0.9082	0.9009	0.9054	0.9155	0.8368	0.8057	0.8158	0.8252
✓		0.9120	0.9057	0.9138	0.9206	0.8382	0.8081	0.8179	0.8247
	✓	0.9136	0.9079	0.9169	0.9204	0.8384	0.8076	0.8200	0.8275
✓	✓	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244

Table 7: Experimental results of decoding global tokens and local tokens in the TransMixer module on the DeepCrack and CamCrack789 datasets.

Methods	DeepCrack				CamCrack789
Methods	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
Avgpooling	0.9147	0.9089	0.9192	0.9140	0.8399	0.8100	0.8227	0.8273
Maxpooling	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244

Table 8: Experimental results of different pooling operations in the Local Refinement Module on the DeepCrack and CamCrack789 Datasets.

Rearrange	DEG	EdgeConv	Gating	DeepCrack				CamCrack789
Rearrange	DEG	EdgeConv	Gating	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
				0.9052	0.8980	0.9042	0.9052	0.8334	0.8008	0.8122	0.8114
✓				0.9074	0.9004	0.9127	0.9130	0.8338	0.8009	0.8079	0.8162
✓	✓			0.9107	0.9044	0.9143	0.9174	0.8376	0.8065	0.8237	0.8252
✓	✓	✓		0.9132	0.9072	0.9125	0.9205	0.8398	0.8092	0.8193	0.8254
✓	✓	✓	✓	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244

Table 9: The ablation experiment results of DEGConv on the DeepCrack and CamCrack789 datasets, including the Rearrange, DEG, Edge Convolution, and Gating operations.

Depth	FLOPs (G)	Params (M)	Memory (MiB)	DeepCrack				CamCrack789
Depth	FLOPs (G)	Params (M)	Memory (MiB)	mIoU	ODS	OIS	F1	mIoU	ODS	OIS	F1
1	2.05	2.54	1190	0.9151	0.9094	0.9197	0.9205	0.8409	0.8115	0.8202	0.8244
2	3.51	4.76	1550	0.9141	0.9084	0.9213	0.9071	0.8381	0.8069	0.8222	0.8208
4	6.42	9.20	2080	0.9126	0.9065	0.9166	0.9173	0.8375	0.8065	0.8237	0.8252
6	9.33	13.63	3818	0.9073	0.9004	0.9127	0.9130	0.8336	0.8005	0.8134	0.8203

Table 10: Experiments on different layer depths on the DeepCrack and CamCrack789 Datasets.

$n$	mIoU	ODS	OIS	F1
9	0.7805	0.7269	0.7467	0.7575
18	0.7814	0.7273	0.7441	0.7631
36	0.7824	0.7281	0.7483	0.7755
90	0.7813	0.7264	0.7455	0.7692
180	0.7774	0.7212	0.7409	0.7592

Table 11: Ablation studies of the number of bins

n

on the Crack500 dataset.

Figure 5: Representative crack images from the Crack500 dataset.

Figure 6: Visual comparison results of multi-scale features before and after SRF module processing. The distinction between cracks and the background is more prominent, and more notably, the crack features in the layer 4 have been significantly enhanced.

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
SCRWKV: Ultra-Compact Structure-Calibrated Vision-RWKV for Topological Crack Segmentation
May 14, 2026Hanxu Zhang, Chen Jia, Hui Liu +3Crack Segmentation Semantic Segmentation
Engineering Research Center of Learning-Based Intelligent System, Ministry of Education, Tianjin University of Technology, Tianjin 300384, China · Key Laboratory of Computer Vision and System, Ministry of Education, Tianjin University of Technology, Tianjin 300384, China
Multi-Task Crack Foundation Model for Engineering-Reliable Crack Representation and Topology Preservation in Civil Infrastructure
Jun 4, 2026Blessing Agyei Kyem, Joshua Kofi Asamoah, Eugene Denteh +1Vision Foundation Model Adaptation Topology-Preserving Image Segmentation
Department of Civil, Construction and Environmental Engineering, North Dakota State University, Fargo, 58102, ND, USA

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

SCRWKV: Ultra-Compact Structure-Calibrated Vision-RWKV for Topological Crack Segmentation

Multi-Task Crack Foundation Model for Engineering-Reliable Crack Representation and Topology Preservation in Civil Infrastructure