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

Conference: CVPR 2026
arXiv: 2603.01361
Code: GitHub
Area: Segmentation / Crack Segmentation
Keywords: Crack segmentation, Hybrid architecture, Mamba attention decoupling, Direction-guided edge convolution, Lightweight and efficient

TL;DR¶

Ours proposes MixerCSeg, which decouples channels into global/local branches by analyzing the implicit attention mechanism of Mamba. These branches are respectively enhanced with Self-Attention and CNN, combined with direction-guided edge gated convolutions. It achieves SOTA performance in crack segmentation with 2.05 GFLOPs and 2.54M parameters.

Background & Motivation¶

Crack segmentation is a critical technology for infrastructure health monitoring, yet it faces challenges such as diverse crack morphologies, uneven texture distribution, and low contrast against backgrounds. Existing architectures have specific shortcomings:

CNN: Strong local feature extraction but insufficient global modeling, making it difficult to handle complex morphologies.
Transformer: Strong global dependency modeling but high computational overhead.
Mamba: Global focus with linear complexity, but its sequential processing mechanism limits global context utilization within a single forward pass.

Existing hybrid models (MambaVision, RestorMixer) involve simple stacking of different architectures without deep analysis of their internal interaction logic. The Key Insight of this paper: Mamba's implicit attention naturally differentiates into global and local channels (identified through \(\Delta_t\) analysis), allowing targeted allocation for CNN, Transformer, and Mamba modules.

Method¶

Overall Architecture¶

MixerCSeg aims to resolve the Key Challenge of crack segmentation: "diverse morphology, low contrast, and the need for lightweight deployment." The structure is an encoder-decoder: the input passes through a Stem layer, followed by multiple TransMixer Blocks to extract multi-scale features \(\{F_1, F_2, F_3, F_4\}\); these features are injected with direction and edge priors via DEGConv, then fused by the SRF module where high-resolution details guide low-resolution semantics, finally reaching the segmentation head for pixel-level crack masks. The Mechanism is "to let CNN, Transformer, and Mamba handle the channels they specialize in" rather than simply stacking them.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Input Crack Image"] --> B["Stem Layer"]
    B --> TM
    subgraph TM["TransMixer Block (Decouple channels via Δt)"]
        direction TB
        M["Mamba Output Y<br/>Sorted by Δt along channels"]
        M -->|"top d·γ global channels"| G["Global Branch<br/>Self-Attention for long-range dependencies"]
        M -->|"remaining d·(1−γ) local channels"| L["Local Branch<br/>Local Refinement for fine details"]
    end
    TM --> F["Multi-scale Features F1–F4"]
    F --> D["DEGConv<br/>Direction Histogram Gated Edge Conv"]
    D --> S["SRF Fusion<br/>High-res details guide low-res semantics"]
    S --> H["Segmentation Head → Pixel-level Mask"]

Key Designs¶

1. TransMixer Block: Channel Decoupling via Mamba \(\Delta_t\)

The Design Motivation stems from the "mindless stacking" issues in previous hybrid architectures. Ours first processes a standard Mamba (Eq.1-2) to obtain output \(Y\), then uses \(\Delta_t\) (the factor controlling the influence of historical tokens on the current token) as a criterion for global/local differentiation. Channels are sorted: those with high \(\Delta_t\) value (fast decay, concentrating more on current tokens/global structure) are selected as the top \(d_g = d \cdot \gamma\) global tokens, while the rest \(d_l = d \cdot (1-\gamma)\) serve as local tokens (default \(\gamma = 0.5\)).

The global branch uses Self-Attention to supplement long-range dependencies, while the local branch utilizes a Local Refinement Module (Norm → Reshape → MaxPool2d → Conv \(1\times 1\) → Sigmoid gating → multiply with original features) for fine-grained details.

2. Direction-guided Edge Gated Convolution (DEGConv)

Cracks are slender structures with explicit orientations, which isotropic convolutions fail to perceive. DEGConv injects direction priors in three steps: (a) Rearrange—Split feature maps into \(N\) non-overlapping local views \(F_i^j \in \mathbb{R}^{C_i \times h_i \times w_i}\) for independent processing; (b) Direction Embedding Generation—Calculate horizontal/vertical gradients using Sobel operators, obtain arc direction \(\theta = \arctan(d_y/d_x)\), construct direction histograms by cell and bin, and compress into a direction embedding vector \(\epsilon \in \mathbb{R}^{C_i}\) via Conv + ReLU + AvgPool; (c) Gated Edge Convolution—\(g = \sigma_2(\text{EdgeConv}(F_i^j + \epsilon))\), \({F_i^j}' = g \odot \text{EdgeConv}(F_i^j)\), where EdgeConv uses \(1 \times k\) and \(k \times 1\) strip convolutions followed by concatenation and depthwise convolution.

3. Spatial Refinement Multi-Level Fusion (SRF)

In direct multi-scale feature concatenation, high-resolution edge details are easily overwhelmed by low-resolution semantics. SRF uses high-resolution features \(F_1'\) to generate a spatial attention map \(\alpha = \sigma_2(\text{Conv}_{1\times1}(F_1'))\), point-wise weighting the upsampled low-resolution features \(F_i'' = \alpha \odot F_i^{up}\).

Loss & Training¶

BCE + Dice Loss, ratio 1:5.
Single NVIDIA A100, 50 epochs, batch size=1.
AdamW optimizer, initial lr=5e-4.
Input size 512×512.
Hyperparameters: \(\gamma=0.5\), cell size=(8,8), bin number \(n=180\) (36 for Crack500 due to smoother curvature and larger width).

Key Experimental Results¶

Main Results¶

Dataset	Metric (mIoU)	MixerCSeg (Ours)	Prev. SOTA	Gain
DeepCrack	mIoU	0.9151	0.9022 (SCSegamba)	+1.43%
CamCrack789	mIoU	0.8409	0.8372 (U-Net)	+0.44%
CrackMap	mIoU	0.8123	0.8094 (SCSegamba)	+0.36%
Crack500	mIoU	0.7824	0.7778 (SCSegamba)	+0.59%
DeepCrack	F1	0.9205	0.9110 (SCSegamba)	+1.04%

Model	FLOPs (G)	Params (M)	Memory (MiB)
MixerCSeg (Ours)	2.05	2.54	1190
SCSegamba	18.16	2.80	2206
RestorMixer	98.71	3.19	10384
MambaVision	642.86	13.57	5222

Ablation Study¶

Config	DeepCrack mIoU	CamCrack mIoU	Description
Baseline (VMamba+Segformer)	0.8826	0.8283	No extra modules
+ TransMixer	0.9016	0.8359	Significant encoder enhancement
+ DEGConv	0.9097	0.8381	Direction/edge modeling
+ SRF	0.9151	0.8409	Multi-level fusion refined

Key Findings¶

MixerCSeg reduces FLOPs by 88.7% compared to SCSegamba while achieving higher mIoU—the efficiency advantage is extremely significant.
TransMixer is more effective than simple stacking methods (MambaVision, RestorMixer), validating that decoupling based on attention characteristics is superior.
\(\gamma = 0.5\) (equal global/local split) is the optimal channel allocation ratio.
Memory usage is only 1190 MiB, making it suitable for edge deployment.

Highlights & Insights¶

Architecture Design from Mechanism Analysis: Instead of intuitive mixing, it identifies channel-level global/local differentiation by analyzing Mamba's \(\Delta_t\) weights.
Direction embeddings introduce crack-specific prior knowledge (Sobel → Direction Histogram → Embedding), enhancing perception of irregular geometries.
Extreme Efficiency: 2.05 GFLOPs + 2.54M parameters, one to two orders of magnitude smaller than most methods with better performance.
SRF fuses features via high-resolution guidance rather than simple concatenation without adding computational cost.

Limitations & Future Work¶

Validated only on crack segmentation; generalizability to general semantic segmentation (e.g., Cityscapes) needs verification.
Spatial block partitioning in DEGConv might cause discontinuities at boundaries, though mitigated by a subsequent EdgeConv layer.
Bin counts in direction histograms require manual tuning and lack an adaptive mechanism.

SCSegamba: Pioneer in using Mamba for crack segmentation, designing structure-aware scanning strategies.
MambaVision: First Mamba-Transformer hybrid vision backbone, but relies on simple stacking.
Insight: Designing architectures from internal model attention mechanisms (rather than intuitive concatenation) is a more principled hybrid strategy.

Rating¶

Novelty: ⭐⭐⭐⭐ (Channel decoupling based on Mamba implicit attention is novel and theoretically grounded)
Experimental Thoroughness: ⭐⭐⭐⭐ (4 datasets, 7 SOTA comparisons, complete ablation and efficiency analysis)
Writing Quality: ⭐⭐⭐⭐ (Clear diagrams, logical derivation from theory to architecture)
Value: ⭐⭐⭐⭐ (Excellent trade-off between efficiency and accuracy for practical applications)