Structure-Field Encoder (SFE)

• SFE is a crack-specific backbone that models global topological continuity by integrating multi-scale detail injection with dynamic state-space propagation.
• It employs an Adaptive Multi-scale Cascaded Modulator (AMCM) to initialize a structure-aware feature field and SCIUs to refine spatial and channel features through geometry-guided token realignment.
• Operating with linear spatial complexity and sub-million parameters, SFE delivers precise crack segmentation performance, achieving an F1 score of 0.8428 on complex datasets.

Searching arXiv for the specified paper and closely related context papers on SCRWKV/RWKV crack segmentation. Structure-Field Encoder (SFE) is the backbone of SCRWKV, the “Ultra-Compact Structure-Calibrated Vision RWKV” proposed for topological crack segmentation. In that architecture, SFE “achieves high-precision modeling via a novel Structure-Field Encoder (SFE) backbone while maintaining linear complexity,” and is designed to “explicitly build global topological continuity of hairline cracks while remaining computationally linear” (Zhang et al., 14 May 2026). Within the full SCRWKV system, SFE is paired with a lightweight Cross-Scale Harmonic Fusion (CSHF) decoder; the complete model has only 1.22M parameters and reports an F1 score of 0.8428 and mIoU of 0.8512 on the TUT dataset (Zhang et al., 14 May 2026).

1. Definition and problem setting

The SFE was introduced in the context of a problem framed as follows: “Achieving pixel-level accurate segmentation of structural cracks across diverse scenarios remains a formidable challenge.” Existing methods are described as facing “significant bottlenecks in balancing crack topology modeling with computational efficiency, often failing to reconcile high segmentation quality with low resource demands” (Zhang et al., 14 May 2026).

Within this formulation, SFE is not a generic vision encoder. It is a crack-specific backbone whose stated function is to model “global topological continuity of hairline cracks” while preserving low computational cost. That emphasis on topology is central: the encoder is constructed so that local texture representation, crack-manifold alignment, and global message passing are all performed inside a linear-complexity framework. This suggests that the design target is not merely boundary sharpness or local contrast enhancement, but continuity-preserving segmentation under “complex textures and severe interference.”

2. Encoder organization within SCRWKV

After “a lightweight patch-embedding and positional encoding,” SFE proceeds in two stages: first, “a one-shot Adaptive Multi-scale Cascaded Modulator (AMCM) ‘initializes’ the feature field by injecting multi-scale local detail and a coarse structural prior”; second, “a stack of $L$ Structure-Calibrated Insight Units (SCIUs) refines that field” (Zhang et al., 14 May 2026).

The SCIU itself is organized into two phases. In “Spatial Mix,” “Geometry-guided Bidirectional Structure Transformation (GBST) realigns tokens along crack manifolds, and Dy-WKV with Dynamic Self-Calibrating Decay (DSCD) performs global, content-aware message passing.” In “Channel Mix,” “the same GBST prior guides an adaptive, gated linear projection for improved channel interactions.” This division gives SFE a layered operational logic: AMCM prepares a structure-aware feature field, and SCIUs iteratively recalibrate that field in both spatial and channel domains.

Component	Role in SFE	Key operations
AMCM	“initializes” the feature field	multi-scale local detail, coarse structural prior
SCIU	refines the field	Spatial Mix with GBST and Dy-WKV+DSCD; Channel Mix with GBST-guided adaptive gating

A concise operational summary is:

$$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$9

3. Adaptive Multi-scale Cascaded Modulator

AMCM is described as the “Structure-Field Initializer.” Given $X_{in}\in\mathbb{R}^{C\times H\times W}$, channels are first split into $X_{p1}$ and $X_{p2}$, followed by a depthwise-gated modulation:

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$

The tensor $\tilde X$ is then sliced into $\{Q_1,Q_2,Q_3\}$, and a cascaded state evolution is applied through

$$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$

with

$$H_1=\Psi_{7}(Q_3), \qquad H_2=\Psi_{9}\bigl(W_{proj}(Q_2)\|\!H_1\bigr).$$

Fusion with $Q_1$ yields

$X_{in}\in\mathbb{R}^{C\times H\times W}$0

where $X_{in}\in\mathbb{R}^{C\times H\times W}$1 is a dilated conv (Zhang et al., 14 May 2026).

AMCM then performs a multi-scale branch aggregation:

$X_{in}\in\mathbb{R}^{C\times H\times W}$2

Grid pooling produces

$X_{in}\in\mathbb{R}^{C\times H\times W}$3

after which $X_{in}\in\mathbb{R}^{C\times H\times W}$4 and $X_{in}\in\mathbb{R}^{C\times H\times W}$5 are inferred via small MLPs on $X_{in}\in\mathbb{R}^{C\times H\times W}$6. With a fixed topology matrix $X_{in}\in\mathbb{R}^{C\times H\times W}$7,

$X_{in}\in\mathbb{R}^{C\times H\times W}$8

The AMCM output is

$X_{in}\in\mathbb{R}^{C\times H\times W}$9

with $X_{p1}$0 a $X_{p1}$1 conv and $X_{p1}$2.

Factually, AMCM “inject[s] multi-scale local detail and a coarse structural prior.” In contextual terms, the sequence of depthwise gating, cascaded large-kernel processing, and topology-modulated grid attention suggests that AMCM is intended to initialize a feature field in which crack texture and coarse continuity cues are already entangled before the recurrent SCIU refinement begins.

4. Structure-Calibrated Insight Unit

SCIU is the core refinement engine of SFE. In its Spatial Mix phase, the first step is GBST. The input $X_{p1}$3 is partitioned along channels into two halves, $X_{p1}$4 and $X_{p1}$5, each subdivided into four directional groups of size $X_{p1}$6. Let $X_{p1}$7 denote a spatial shift by offset $X_{p1}$8. The outward and inward streams are defined as

$X_{p1}$9

with $X_{p2}$0, and

$X_{p2}$1

This is the mechanism by which tokens are said to be “realign[ed] along crack manifolds” (Zhang et al., 14 May 2026).

The GBST-transformed feature is blended with the original input using

$X_{p2}$2

Linear projections then generate receptance, key, and value:

$X_{p2}$3

DSCD introduces a dynamic decay vector

$X_{p2}$4

with $X_{p2}$5. After flattening $X_{p2}$6, a structure-aware score is defined for each query-position $X_{p2}$7 and key-position $X_{p2}$8:

$X_{p2}$9

This yields weights

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$0

and aggregation

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$1

The spatial output is

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$2

The Channel Mix phase introduces

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$3

then, for each channel-group $$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$4,

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$5

The channel output is

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$6

and the SCIU output is

$$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$7

In descriptive terms, SCIU couples geometry-guided token rearrangement with dynamic state-space-style propagation. The paper’s formulation makes the intended separation explicit: Spatial Mix handles crack-aligned global propagation, whereas Channel Mix uses the same structural prior to improve channel interactions through adaptive gating.

5. Mathematical scaling and parameter budget

The complexity discussion in the source contains two complementary descriptions. At the architectural overview level, “all operations are either local convolutions or linear-complexity state-space–style message passing, so SFE’s total cost grows only $$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$8, in contrast to $$\tilde X = X_{in} \odot [\,X_{p1}\|\!D_{3}(X_{p2})\,].$$9 for a standard vision-transformer.” In the detailed FLOP accounting, the total per image is written as

$\tilde X$0

and the encoder is described as “maintaining strictly linear $\tilde X$1 complexity” because it removes the quadratic dependence on spatial token count characteristic of self-attention (Zhang et al., 14 May 2026).

The parameter analysis is given explicitly. For AMCM:

• pointwise convs $\tilde X$2: $\tilde X$3,
• depthwise large kernels $\tilde X$4 with $\tilde X$5: $\tilde X$6,
• small grid MLPs for $\tilde X$7: $\tilde X$8,

leading to “$\tilde X$9, since $\{Q_1,Q_2,Q_3\}$0 is constant.”

For each SCIU:

• $\{Q_1,Q_2,Q_3\}$1: $\{Q_1,Q_2,Q_3\}$2,
• GBST and DSCD have no extra parameters beyond $\{Q_1,Q_2,Q_3\}$3,

so the per-block count is “$\{Q_1,Q_2,Q_3\}$4,” and for $\{Q_1,Q_2,Q_3\}$5 blocks the contribution is “$\{Q_1,Q_2,Q_3\}$6.”

The resulting encoder budget is summarized as

$\{Q_1,Q_2,Q_3\}$7

“plus a small $\{Q_1,Q_2,Q_3\}$8 term.” The concrete examples given are:

• “For $\{Q_1,Q_2,Q_3\}$9, $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$0; plus AMCM’s $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$1 total.”
• “In practice we set $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$2 and get $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$3.”

The comparison baseline is also explicit: “A standard ViT encoder layer costs $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$4 in self-attention plus $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$5,” whereas “a ResNet conv block costs $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$6.” The stated conclusion is that “SFE sits between pure CNN (slightly larger effective kernel) and Transformer (no quadratic term).”

6. Segmentation role, reported performance, and interpretive issues

In the complete SCRWKV system, SFE is combined with “a lightweight Cross-Scale Harmonic Fusion (CSHF) decoder to achieve precise feature aggregation.” Systematic evaluations are reported on “multiple benchmarks characterized by complex textures and severe interference,” and SCRWKV is said to “significantly outperform SOTA methods.” On the TUT dataset, the full model achieves “an F1 score of 0.8428 and mIoU of 0.8512,” while retaining “only 1.22M parameters,” which the paper presents as evidence of “robust potential for efficient real-world deployment” (Zhang et al., 14 May 2026).

For SFE specifically, the summary characterization is that it “achieves global crack-topology modeling by mixing multi-scale local detail (AMCM) with geometry-aware, dynamic state-space propagation (SCIUs), all within strictly linear complexity and sub-million parameter counts.” A common point of confusion is whether the module should be regarded as primarily convolutional or primarily transformer-like. The source describes it as neither in a strict sense: its operations are “either local convolutions or linear-complexity state-space–style message passing,” and its computational profile is positioned “between pure CNN” and “Transformer.” Another possible misunderstanding is to equate “linear complexity” with the absence of $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$7 terms; the detailed complexity statement instead indicates linearity with respect to spatial extent, specifically by eliminating the quadratic $$\Psi_{\kappa}(h)=W_{p2}\Bigl(\bigl[D_{\kappa}\bigl(\delta(W_{p1}(h))\bigr)\bigr]\odot W_{p1}(h)\Bigr),$$8 self-attention term.

A plausible implication is that SFE’s principal contribution is architectural calibration rather than scale alone: AMCM establishes a structure-aware feature field, and SCIU repeatedly imposes geometry-guided spatial and channel refinement. In the setting of fine-grained crack segmentation, that organization is consistent with the paper’s emphasis on “topologically continuous crack segmentation on resource-constrained devices.”