PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2603.01361v1cs.CVcs.AI

TL;DR

Pixel-level crack segmentation must represent fine textures, thin structures, and complex crack patterns despite limitations in existing CNN-, Transformer-, and Mamba-based encoders. MixerCSeg addresses this with a coordinated TransMixer encoder, DEGConv edge refinement, and SRF multi-level fusion. Across multiple crack segmentation benchmarks, it reports state-of-the-art performance with 2.05 GFLOPs and 2.54 M parameters.

  • Problem

    Existing CNN-, Transformer-, and Mamba-based crack segmentation models each capture only part of the spatial or structural information needed for complex crack patterns.

  • Method

    MixerCSeg coordinates CNN, Transformer, and Mamba pathways in TransMixer, then uses DEGConv for direction-aware edge modeling and SRF for multi-level feature refinement.

  • Results

    MixerCSeg achieves state-of-the-art performance across multiple crack segmentation benchmarks with 2.05 GFLOPs and 2.54 M parameters.

  • Takeaways & Limitations

    The reported results support an efficient hybrid architecture that combines local, global, and contextual-flow modeling for pixel-level crack segmentation.

Abstract

from arXiv · show

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.

1. Introduction

Road crack segmentation is difficult because cracks vary morphologically, have uneven textures, and often contrast weakly with backgrounds. MixerCSeg addresses limitations across CNN, Transformer, and Mamba approaches with a coordinated hybrid encoder and lightweight refinement modules.

  • 1. Introduction: Pixel-level crack segmentation remains challenging because cracks exhibit morphological diversity, uneven texture distribution, and low contrast against backgrounds.These conditions motivate architectures that model both fine details and broader context.
  • 1. Introduction: CNN-based models efficiently extract local features but struggle with long-range dependencies, while Mamba-based processing can limit global-context capture in one forward pass.The supplied discussion contrasts local receptive-field limitations with progressive Mamba processing.
  • 1. Introduction: DEGConv enhances direction-aware edge modeling, and SRF refines multi-scale spatial details with minimal computational cost.DEGConv uses spatial block processing and directional priors, while SRF provides multi-level refinement.
  • 1. Introduction: MixerCSeg combines CNN, Transformer, and Mamba pathways to capture local textures, global dependencies, and contextual flow for crack segmentation.TransMixer assigns distinct roles to the three pathways rather than simply stacking modules.
  • 1. Introduction: 2.05 GFLOPs and 2.54 M parameters accompany state-of-the-art performance across multiple crack segmentation benchmarks.The reported design emphasizes both segmentation performance and computational efficiency.

2. Related Work

Crack segmentation research spans CNN, Transformer, and Mamba models, while hybrid architectures increasingly combine their complementary capabilities. MixerCSeg targets a remaining gap: designing coordinated interaction among CNN, Transformer, and Mamba components rather than merely stacking them.

  • 2. Related Work: CNN, Transformer, and Mamba methods respectively emphasize efficient local features, long-range dependencies, and linear-complexity global processing for crack segmentation.The related work describes representative methods and their architectural advantages.
  • 2. Related Work: Mamba-Transformer-CNN hybrids remain exploratory in visual tasks, although prior hybrid models have shown benefits in vision, restoration, forecasting, and video understanding.Examples include MambaVision, RestorMixer, MambaFormer, and VAMBA.
  • 2. Related Work: MixerCSeg’s overview connects TransMixer blocks and downsampling for multi-scale features, then applies DEGConv before decoding to enhance crack texture details.The architecture also includes directional priors and spatial block processing within DEGConv.
  • 2. Related Work: Existing hybrid models often stack Mamba and Transformer modules without refining how the architectures collaboratively interact.MixerCSeg is motivated by analyzing Mamba’s hidden attention to assign complementary roles across architectures.

3. Methodology

MixerCSeg analyzes Mamba’s hidden attention to coordinate global and local feature processing, then combines directional edge modeling with spatially refined multi-level fusion for crack segmentation.

  • 3.1. Preliminary: Hidden Attention of Mamba: Mamba’s state-space computation assigns historical-token influence through decay and update factors, revealing an attention-like mechanism for modeling token dependencies.The decay factor remains between 0 and 1, while Δt determines which tokens influence future tokens.
  • 3.2. Overall Architecture: MixerCSeg processes TransMixer outputs into multi-scale features, uses DEGConv for texture and topological cues, and applies SRF for efficient multi-level feature fusion.The architecture produces pixel-level crack predictions after these sequential feature-processing stages.
  • 3.3. The Design of TransMixer: TransMixer sorts channels by Δt into global and local tokens, applying self-attention to global representations and Local Refinement to local representations.This design replaces simple stacking of Mamba and Transformer modules with an explicit division of their complementary processing roles.
  • 3.4. Direction-guided Edge Gated Convolution: DEGConv partitions feature maps into local views, derives Sobel-based directional embeddings, and uses gated EdgeConv processing to model intersecting and branching crack structures.Direction histograms aggregate pixel orientations within spatial cells before forming a direction embedding.
  • 3.5. Spatial Refinement Multi-Level Feature Fusion: SRF addresses boundary misalignment and underuse of high-resolution details by refining spatially inconsistent multi-level features before segmentation.Its motivation is to improve fine-grained spatial information without increasing complexity.

4. Experiments

MixerCSeg achieves stronger crack-segmentation performance than advanced baselines across multiple datasets while maintaining low computational cost. Ablations attribute gains to TransMixer, DEGConv, and SRF, with directional and hyperparameter studies supporting the design choices.

  • Quantitative Evaluation: 1.43% higher mIoU and 1.04% higher F1-score than SCSegamba on DeepCrack establish MixerCSeg’s benchmark advantage.It also exceeds MambaVision by 1.78% in mIoU and 4.61% in F1-score on DeepCrack.
  • Quantitative Evaluation: 2.05 GFLOPs and 2.54 M parameters make MixerCSeg substantially cheaper than competing lightweight and hybrid models.Against SCSegMamba, it reduces parameters by 9.3%, computational cost by 88.7%, and memory usage by 1016 MiB.
  • Qualitative Evaluation: Visual comparisons evaluate MixerCSeg across challenging environments containing variable crack sizes, diverse morphologies, and background noise.The supplied visualization passages identify the comparison scope but do not provide specific visual outcomes.
  • Ablation Study: Component ablations show that TransMixer, DEGConv, and SRF each improve segmentation, with their combination achieving state-of-the-art performance.DEGConv adds only 0.08 GFLOPs and 0.14 M parameters, while SRF reduces decoder cost and memory.
  • Ablation Study: The default γ = 0.5 and 8 × 8 DEGConv cell size outperform tested alternatives, while excessively large directional-bin counts reduce performance.Smaller cells are more sensitive to background noise, and overly fine binning can leave bins with only one value.

5. Conclusion

MixerCSeg combines Mamba, Transformer, and CNN strengths through TransMixer, while DEGConv and SRF improve structural and pixel-level refinement. Experiments show state-of-the-art crack segmentation with lower computational requirements.

  • 5. Conclusion: MixerCSeg combines Mamba, Transformer, and CNN pathways to capture multi-scale crack features efficiently.TransMixer separates local and global tokens, DEGConv uses directional priors for texture and morphology, and SRF refines low-resolution features with high-resolution information.
  • 5. Conclusion: Extensive benchmark experiments report state-of-the-art crack segmentation while requiring lower computational resources.

7. Additional Ablation Experiments

Ablations show that TransMixer’s global and local token pathways, DEGConv components, block depth, and bin count each affect crack-segmentation performance and efficiency. The experiments support depth 1 and n = 36 as selected configurations while identifying measurable component contributions.

  • 7.1. Ablation studies of the TransMixer Module: 0.76% and 0.49% performance declines on DeepCrack and CamCrack789, respectively, follow removal of TransMixer, while removing either token type also reduces performance.The ablation attributes meaningful roles to both global and local tokens.
  • 7.1. Ablation studies of the TransMixer Module: Maxpooling enhances salient crack features within local regions more effectively than avgpooling.This comparison evaluates pooling choices in the Local Refinement Module.
  • 7.2. Ablation studies of the DEGConv Module: 0.36% and 0.46% mIoU gains on DeepCrack and CamCrack789, respectively, come from DEG, while Rearrange adds 0.24% and 0.05%.The results report contributions from each DEGConv component and describe Rearrange as structuring features for later operations.
  • 7.3. Ablation studies of TransMixer Block Depth: Depth 1 is selected for each TransMixer block because deeper configurations increase optimization difficulty, degrade performance, and raise model complexity.The authors connect the selected depth to accuracy, complexity control, and edge-device feasibility.
  • 7.4. Ablation study of the Number of Bins: n = 36 gives the best Crack500 performance among the tested bin counts 9, 18, 36, 90, and 180.The accompanying analysis links this setting to Crack500’s dominant crack morphology and noise characteristics.

8. Qualitative Analysis of SRF Module

SRF improves multi-scale feature representations by strengthening crack-background discrimination and aligning high-level semantics with high-resolution details. It also reduces decoder resource usage while improving mIoU on two crack-segmentation benchmarks.

  • 8. Qualitative Analysis of SRF Module: 89.3% lower computational cost, 33.8% fewer parameters, and 67.2% lower GPU memory usage result when SRF replaces the SegFormer decoder.The replacement also improves mIoU by 0.59% on DeepCrack and 0.32% on CamCrack789.
  • 8. Qualitative Analysis of SRF Module: SRF makes crack regions and backgrounds more semantically distinguishable across multi-scale feature maps, with stronger crack activation and clearer boundaries.The visual analysis reports especially enhanced crack features in layer 4.
  • 8. Qualitative Analysis of SRF Module: SRF improves alignment between refined features and high-resolution details, alleviating insufficient fusion of high-level semantics and low-level decoder information.The resulting features provide more discriminative cues for the segmentation head.
Loading 2603.01361v1…