Ultra-Compact SCRWKV for Crack Segmentation

• The paper introduces SCRWKV, a novel vision architecture that integrates a structure-field encoder and cross-scale harmonic fusion to achieve pixel-accurate crack extraction.
• It employs adaptive multi-scale cascaded modulation, geometry-guided bidirectional token interaction, and dynamic noise-aware decay to tackle sparse, noisy crack pixels in challenging scenes.
• Empirical results across multiple datasets demonstrate high F1 and mIoU scores while maintaining model compactness and efficient deployment.

Ultra-Compact Structure-Calibrated Vision RWKV (SCRWKV) is a vision architecture for topological crack segmentation that targets pixel-accurate crack extraction in challenging structural scenes while maintaining a highly compact computational profile. Introduced in "SCRWKV: Ultra-Compact Structure-Calibrated Vision-RWKV for Topological Crack Segmentation" (Zhang et al., 14 May 2026), it is designed for pavement, concrete, tiles, runways, and metal surfaces, where cracks are thin, elongated, often discontinuous, curvilinear, arbitrarily oriented, and embedded in heavy background interference. The model combines a Structure-Field Encoder (SFE), a Structure-Calibrated Insight Unit (SCIU), and a Cross-Scale Harmonic Fusion (CSHF) decoder to reconcile topology modeling with linear-complexity global aggregation.

1. Problem setting and design objective

The central problem addressed by SCRWKV is pixel-accurate crack segmentation in highly challenging real-world structural scenes (Zhang et al., 14 May 2026). The paper identifies several domain-specific difficulties. Cracks are non-rigid, fragmented, self-similar at multiple scales, and spatially sparse. They are also topologically complex, with bifurcations, winding paths, and varying widths, while crack pixels typically occupy only a tiny fraction of the image.

Background interference is severe. The paper lists road texture, oil stains, shadows, pavement particles, and illumination changes as common sources of confusion. Under these conditions, local edge sensitivity alone is insufficient: a crack may stretch across the full image, so preserving global continuity is necessary to avoid broken or disconnected masks.

SCRWKV is positioned against three established modeling regimes. CNNs have local receptive fields and are therefore strong at local textures but weak at long-range crack continuity. Transformers model global context but incur quadratic complexity. Mamba/SSM methods provide linear complexity, but typically flatten images into 1D scans, which can damage 2D crack continuity. Prior Vision-RWKV methods inherit rigid spatial shifting and static decay, which the paper argues are poorly matched to irregular crack geometry. The design objective is therefore explicit: to jointly achieve strong topology modeling and low computational cost suitable for edge deployment.

2. Overall architecture

SCRWKV consists of three main stages: patch embedding with positional encoding, the Structure-Field Encoder backbone, and the Cross-Scale Harmonic Fusion decoder (Zhang et al., 14 May 2026). The backbone produces multi-scale features $\{F_1,\dots,F_4\}$, which are fused by the decoder into a binary crack mask.

The SFE is intended to model crack topology as a "structure field." It contains an initial Adaptive Multi-scale Cascaded Modulator (AMCM) followed by multiple Structure-Calibrated Insight Unit blocks. The stated goals of the backbone are to strengthen local morphology, preserve global continuity, suppress noise, and keep the model extremely compact.

The architecture’s main conceptual combination is stated explicitly in the paper: structure-field initialization and enrichment, geometry-guided bidirectional token interaction, dynamic noise-aware decay in RWKV recurrence, and lightweight cross-scale decoder fusion. Relative to standard Vision-RWKV, the paper identifies two specific replacements: GBST instead of rigid Q-Shift, and DSCD for content-adaptive decay within Dy-WKV. This establishes SCRWKV as a topology-aware RWKV-style vision network rather than a direct adaptation of standard RWKV blocks.

3. Structure-Field Encoder: AMCM and SCIU

Within the SFE, AMCM functions both as a standalone structure-field initializer and as a local-detail enhancer inside SCIU (Zhang et al., 14 May 2026). Given input $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$, the module first splits the tensor into two parts,

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$

then applies depthwise convolution, concatenation, and Hadamard gating: $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$ The gated feature is further processed by cascaded multi-kernel extraction and parallel multi-scale branches with $\mathcal{K}=\{5,7,9,11\}$, after which a grid-based global calibration stage constructs spatial attention through a learnable topology matrix $\mathbf{S}$. The paper’s stated intuition is that AMCM captures fine cracks via large-kernel local branches and global structural cues via grid attention.

SCIU is the encoder’s core computational block. It combines GBST, Dy-WKV with DSCD, and dynamic channel mixing. GBST is introduced to address the weakness of fixed Q-Shift by explicitly modeling bidirectional geometric propagation. Its directional mapping is defined as

$$\Phi(\mathbf{x}, \Delta) = \sum_{k=1}^{4} \mathcal{T}_{\Delta_k}(\mathbf{x}^{(k)}), \qquad \Delta \in \{\uparrow,\downarrow,\leftarrow,\rightarrow\},$$

and the bidirectional transformation is

$$\text{GBST}(\mathbf{X}_{in}) = \left[ \Phi(\mathbf{x}_{out}, \Delta^{out}) \,;\, \Phi(\mathbf{x}_{in}, \Delta^{in}) \right],$$

with $\Delta^{in} = -\Delta^{out}$. The appendix pseudocode described in the summary states that channels are split into two groups, with an outward diffusion stream that expands signals and an inward convergence stream that preserves the skeleton. The paper interprets this as alignment with curvilinear crack manifolds and preservation of continuity across bends and bifurcations.

Before recurrence, SCIU interpolates between the input and GBST output using a learnable gate: $$\tilde{\mathbf{x}}_c = \mu_c \odot \mathbf{X}_{in} + (1 - \mu_c) \odot \text{GBST}(\mathbf{X}_{in}).$$ From these aligned features it derives receptance $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$0, key $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$1, and value $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$2. The Dynamic Self-Calibrating Decay is then introduced as a content-dependent alternative to the static decay vector of standard RWKV: $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$3 The paper states that this matters because crack pixels are sparse, background noise dominates, and adaptive decay suppresses noisy tokens while reinforcing structure-dominant signals.

The Dy-WKV recurrence defines a structure-aware attention score

$$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$4

followed by weighted bidirectional aggregation and

$$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$5

In the channel branch, dynamic gating is used through

$$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$6

and a GBST-guided mixture

$$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$7

with output projection

$$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$8

where $$\mathbf{X}_{in} \in \mathbb{R}^{C \times H \times W}$$9 is Squared-ReLU. The paper’s interpretation is that this branch preserves richer feature interactions while remaining lightweight.

4. Cross-scale decoding and learning objective

The Cross-Scale Harmonic Fusion decoder is the segmentation head that combines encoder features from different scales and is intended to resolve the semantic gap between shallow, boundary-sensitive features and deep, semantic features (Zhang et al., 14 May 2026). Backbone features $$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$0 are first projected to a common channel dimension and upsampled,

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$1

then augmented with a learnable scale embedding,

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$2

Scale attention is computed as

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$3

and the harmonic fusion itself is

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$4

The final segmentation output is

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$5

The paper emphasizes that CSHF does not merely concatenate features; it performs scale-aware, gated, harmonic blending to improve boundary precision and semantic consistency.

The training objective is

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$6

with

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$7

and

$$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$8

The best ratio is reported as $$\mathbf{X}_{in} \rightarrow \{\mathbf{X}_{p1}, \mathbf{X}_{p2}\},$$9, which the paper attributes to a balance between BCE for sharp pixel supervision and Dice for structural overlap and class imbalance.

5. Empirical performance, compactness, and deployment

The paper evaluates SCRWKV on four datasets: Crack500, DeepCrack, CrackMap, and TUT (Zhang et al., 14 May 2026). Reported metrics are ODS, OIS, Precision, Recall, F1, and mIoU.

Dataset	Key reported result	SCRWKV
Crack500	F1 / mIoU	0.7552 / 0.7787
DeepCrack	F1 / mIoU	0.9363 / 0.9289
CrackMap	F1 / mIoU	0.7740 / 0.8106
TUT	F1 / mIoU	0.8428 / 0.8512

On TUT, which is the headline result emphasized in the abstract, SCRWKV achieves ODS $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$0, OIS $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$1, Precision $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$2, Recall $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$3, F1 $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$4, and mIoU $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$5. On DeepCrack, the paper states that it outperforms the second-best method by $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$6 F1 and $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$7 mIoU. The compared methods include RIND, SFIAN, CTCrackSeg, DTrCNet, Crackmer, SimCrack, MambaIR, CSMamba, PlainMamba, and SCSegamba. The reported interpretation is that SCRWKV is especially strong on DeepCrack, CrackMap, and TUT, while on Crack500 it remains highly competitive and takes ODS and OIS.

The efficiency profile is a central part of the contribution. SCRWKV is reported to have 1.22M parameters, 22.78 GFLOPs at $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$8, and 28 MB model size. In the complexity comparison table, the paper states that SCRWKV has the smallest parameter count and the best model size among the listed methods. Although Crackmer has lower FLOPs at 14.94G, the paper argues that SCRWKV’s slightly higher FLOPs are a worthwhile trade-off for much better accuracy and drastically fewer parameters.

The deployment study uses a UAV-based setting with a DJI Mini SE, a laptop server with Intel Core i7-14650, Ubuntu 22.04, a 30 fps video stream, and input resized to $$\tilde{\mathbf{X}} = \mathbf{X}_{in} \odot \left( \mathbf{X}_{p1} \oplus \mathcal{D}_{3}(\mathbf{X}_{p2}) \right).$$9. In this setup, SCRWKV runs at 46 FPS with 0.0216 s/frame. The paper notes that this is nearly as fast as the fastest competitor, DTrCNet at 47 FPS and 0.0213 s/frame, while being far smaller and, according to the reported benchmarks, more accurate.

6. Ablations, limitations, and significance

The ablation studies test AMCM, GBST, DSCD, and their combinations, and the full model reaches F1 $\mathcal{K}=\{5,7,9,11\}$0, mIoU $\mathcal{K}=\{5,7,9,11\}$1, 1.22M parameters, and 22.78G FLOPs (Zhang et al., 14 May 2026). The reported conclusions are that AMCM improves local morphological modeling, GBST is crucial for preserving topology, DSCD helps suppress noise, and the combination of GBST and DSCD is especially efficient. The paper interprets the three modules as complementary rather than redundant.

Further ablations evaluate CSHF against UNet head, Ham head, SegFormer head, and MFS head; GBST against bi_ParaSnake, DiagSnake, bi_DiagSnake, SASS, Omishift, and Q-Shift; SCIU depth at 2, 4, 8, and 16 layers; patch sizes 4, 8, 16, and 32; and Dice:BCE ratios from 1:5 to 5:1 as well as single-loss baselines. The best reported setting is a 4-layer SCIU stack, patch size 4, and loss ratio 3:1. The paper explains that larger patch sizes hurt fine crack detail and cause mixed-pixel ambiguity, while deeper models increase redundancy and do not improve performance.

The paper also states several limitations. The formal complexity derivation of each module is not fully expanded in the excerpt. The exact fully typeset aggregation formula for RWKV/Dy-WKV is incomplete in the provided text. The dataset focus remains limited to crack segmentation, and broader generalization to other thin-structure segmentation tasks is suggested as future work. In addition, the model is slightly more computationally expensive in FLOPs than the fastest lightweight baseline in some settings, even though it is far smaller in parameters.

Within those boundaries, SCRWKV’s significance lies in its claim that crack segmentation benefits from treating cracks as a geometry-preserving field rather than only as local texture patterns or flattened sequences. This suggests a broader methodological lesson for irregular, sparse, curvilinear targets: topology-aware recurrence, dynamic decay, and scale-aware fusion can be combined without abandoning compactness. The code is reported as available at the project repository linked in the paper (Zhang et al., 14 May 2026).