Technical Summary: SCRWKV for Topological Crack Segmentation

Problem Statement

Achieving pixel-level accurate segmentation of structural cracks in diverse real-world scenarios remains a formidable challenge. Existing methods face a critical bottleneck in balancing crack topology modeling with computational efficiency.

• CNN-based methods (e.g., SFIAN, ECSNet) utilize local inductive biases but suffer from limited receptive fields, often causing segmentation fractures in long-range cracks.
• Transformer-based approaches (e.g., DTrCNet, MFAFNet) capture global dependencies but incur prohibitive quadratic computational complexity, hindering real-time application on resource-constrained edge devices.
• Selective State Space Models (SSMs) like Mamba offer linear complexity but typically rely on predefined fixed-path scanning to flatten 2D images into 1D sequences. This process disrupts the spatial contiguity of curved, bifurcated, and arbitrarily oriented cracks, leading to feature fragmentation.
• Existing RWKV models (e.g., Vision-RWKV) introduce bidirectional operators but often rely on rigid spatial interaction mechanisms (like standard Q-Shift) and static decay. These fail to capture the geometric deformations of cracks and lack dynamic mechanisms to suppress noise propagation in complex backgrounds where cracks occupy only a minority of pixels.

Methodology

The authors propose SCRWKV (Ultra-Compact Structure-Calibrated Vision-RWKV), a network designed to achieve high-precision modeling via linear complexity while explicitly calibrating for structural topology. The architecture consists of a Structure-Field Encoder (SFE) backbone and a Cross-Scale Harmonic Fusion (CSHF) decoder.

1. Structure-Field Encoder (SFE)

The SFE serves as the backbone for hierarchical feature extraction, initialized by a standalone Adaptive Multi-scale Cascaded Modulator (AMCM) and composed of stacked Structure-Calibrated Insight Units (SCIU).

• Adaptive Multi-scale Cascaded Modulator (AMCM): Acts as a structure-field initializer. It employs a divide-and-conquer strategy using parallel gating and a hierarchical cascade of multi-scale large kernels (5×5, 7×7, 9×9, 11×11). This module pre-modulates discrete tokens into an incipient field enriched with preliminary geometric correlations, compensating for the local granularity limitations of linear attention.
• Structure-Calibrated Insight Unit (SCIU): The core building block integrating three mechanisms:
  • Geometry-guided Bidirectional Structure Transformation (GBST): Replaces fixed neighborhood offsets with a mechanism that simulates bidirectional stress field propagation. It partitions feature channels into counter-propagating streams: an Outward-shift Group to diffuse signals (capturing fracture width) and an Inward-shift Group to converge features (preserving the topological skeleton). This explicitly establishes long-range topological dependencies across the image.
  • Dynamic Self-Calibrating Decay (DSCD) in Dy-WKV: Addresses the issue of content-agnostic aggregation. Unlike standard RWKV which treats all spatial tokens equally, Dy-WKV incorporates DSCD to dynamically generate adaptive decay factors. This allows the model to selectively suppress noise-related tokens (e.g., shadows, texture) while amplifying structure-dominant signals during recursive propagation.
  • Spatial and Channel Mixing: The SCIU utilizes these mechanisms to achieve holistic modeling, capturing fine morphological cues while filtering background interference.

2. Cross-Scale Harmonic Fusion (CSHF) Decoder

To bridge the semantic gap between high-level contexts and low-level boundaries, the CSHF decoder employs a scale-signature binding mechanism. It aligns multi-level features via content-aware DySample, injects learnable scale embeddings, and synthesizes a harmonized representation through a scale-aware attention mechanism. This dynamically orchestrates multi-level features into a coherent state for precise binary segmentation.

Key Contributions

1. Ultra-Compact Architecture: The proposal of SCRWKV, a network with only 1.22M parameters, which synergizes global linear modeling with structural awareness. It redefines the efficiency-performance trade-off, delivering SOTA precision with minimal overhead.
2. Novel Structural Components:
   • Development of the SFE backbone featuring the SCIU block, where GBST captures topological continuity and AMCM refines fine morphological cues.
   • Introduction of DSCD for adaptive noise filtering within the Dy-WKV module.
   • Design of the CSHF decoder to effectively bridge semantic gaps via dynamic multi-level fusion.
3. Robust Real-World Performance: Experimental validation demonstrating that extreme compactness does not compromise performance, ensuring applicability in industrial environments with complex textures and severe interference.

Experimental Results

The model was evaluated on four public datasets: Crack500, DeepCrack, CrackMap, and TUT.

• Performance: SCRWKV significantly outperformed 10 SOTA methods (including RIND, SFIAN, MambaIR, CSMamba, and SCSegamba) across all datasets.
  • On the TUT dataset (characterized by complex interference), SCRWKV achieved an F1 score of 0.8428 and an mIoU of 0.8512.
  • On DeepCrack, it achieved an F1 of 0.9363 and mIoU of 0.9289, outperforming the second-best method by margins of 2.53% and 2.67% respectively.
• Efficiency: With only 1.22M parameters and a model size of 28MB, SCRWKV is significantly lighter than competitors (e.g., 56.43% fewer parameters than the second-best).
• Inference Speed: In a real-world UAV deployment scenario, the model achieved an inference time of 0.0216 seconds per frame (46 FPS), comparable to the fastest competitors while maintaining superior topological continuity.
• Ablation Studies:
  • Removing GBST caused a substantial drop in performance, confirming its role in preserving global topology.
  • The 4-layer SCIU configuration was found to be optimal, balancing depth and redundancy.
  • A patch size of 4 was determined optimal for preserving high-frequency crack details.

Significance and Claims

The paper claims that SCRWKV successfully overcomes the bottleneck of balancing topological modeling capability with computational efficiency. By integrating the SFE and CSHF, the framework resolves long-range dependency fragmentation and environmental noise interference without the heavy parameter overhead of Transformers or the topological rigidity of standard SSMs.

The authors position SCRWKV as a premier solution for Structural Health Monitoring (SHM) on resource-constrained edge devices (such as UAVs). The work demonstrates that it is possible to achieve high-precision, real-time crack segmentation in dynamic, noisy environments using an ultra-compact model, thereby facilitating low-cost and automated inspections of civil infrastructure. The code is made available to support further research in this domain.