Strip Receptive Field Module (SRFM)

• SRFM is a neural network building block that uses strip-shaped receptive fields to model long-range, anisotropic context.
• It replaces traditional NxN convolutions with horizontal and vertical strip pooling, reducing parameters while capturing geometric features.
• Integrating SRFM in segmentation and detection backbones has led to state-of-the-art results on benchmarks like ADE20K and RDD2022.

The Strip Receptive Field Module (SRFM) is a neural network building block that achieves efficient and expressive long-range context modeling by constructing anisotropic, strip-shaped receptive fields. SRFM enables convolutional architectures to aggregate information over long horizontal and vertical axes at minimal parameter and computational cost, making it particularly effective for dense prediction tasks that require fine-grained, geometric, or shape-sensitive scene understanding. The SRFM family includes implementations in both semantic segmentation and object detection backbones, having demonstrated state-of-the-art performance on large-scale benchmarks as well as specialized applications such as road damage detection (Hou et al., 2020, Lin et al., 17 Oct 2025).

1. Architectural Design and Variants

SRFM architectures share a foundational approach: they replace conventional NxN square pooling or convolutional kernels with orthogonal strip operators—typically 1xN and Nx1 shapes—on feature tensors $X \in \mathbb{R}^{C \times H \times W}$.

• In the original "Strip Pooling Module" (SPM) (Hou et al., 2020), the input features undergo two parallel strip pooling paths:
  • Horizontal Strip Pooling (HSP): Averages each row across all columns to produce $Y^h \in \mathbb{R}^{C \times H \times 1}$.
  • Vertical Strip Pooling (VSP): Averages each column across all rows to output $Y^v \in \mathbb{R}^{C \times 1 \times W}$.
  • Each path is processed by a 1D convolution (1x3 for HSP, 3x1 for VSP).
  • The outputs are broadcast and summed: $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$.
  • A channel-reweighting block (1x1x1 conv + sigmoid) produces multiplicative attention, which is applied to $X$:

  $Z = X \odot \sigma(f(Y))$ - The output is element-wise multiplied with the input and passed onward, optionally as a residual branch.

• In object detection (e.g., "StripRFNet" (Lin et al., 17 Oct 2025)), SRFM modules:

  • Use a sequence of depthwise $k \times k$ conv for local detail, a horizontal $1 \times K$ conv, followed by a vertical $K \times 1$ conv, then a $1 \times 1$ conv for channel mixing and attention.
  • In parallel, extract global context using 1xW (horizontal) and Hx1 (vertical) average pools, followed by small strip convolutions and broadcast fusion.
  • The outputs are always recombined with the input via a residual sum, ensuring gradient flow and feature reuse.

A typical implementation (as in (Lin et al., 17 Oct 2025)) demonstrates this multi-branch approach:

$X$3

Typical kernel choices are $Y^h \in \mathbb{R}^{C \times H \times 1}$0, $Y^h \in \mathbb{R}^{C \times H \times 1}$1, $Y^h \in \mathbb{R}^{C \times H \times 1}$2 or $Y^h \in \mathbb{R}^{C \times H \times 1}$3, with global strip pooling spanning full width/height for global context.

2. Mathematical Characterization of Receptive Field

The strip pooling and strip convolution operations yield highly anisotropic receptive fields:

• For strip pooling, e.g., horizontal pooling:

  $Y^h \in \mathbb{R}^{C \times H \times 1}$4

• Applying a 1xK followed by a Kx1 convolution produces an effective $Y^h \in \mathbb{R}^{C \times H \times 1}$5 receptive field, but at $Y^h \in \mathbb{R}^{C \times H \times 1}$6 parameter cost per channel versus $Y^h \in \mathbb{R}^{C \times H \times 1}$7 for a dense $Y^h \in \mathbb{R}^{C \times H \times 1}$8 kernel.

  $Y^h \in \mathbb{R}^{C \times H \times 1}$9

  $Y^v \in \mathbb{R}^{C \times 1 \times W}$0

  For $Y^v \in \mathbb{R}^{C \times 1 \times W}$1, $Y^v \in \mathbb{R}^{C \times 1 \times W}$2, $Y^v \in \mathbb{R}^{C \times 1 \times W}$3, $Y^v \in \mathbb{R}^{C \times 1 \times W}$4.

Stacking $Y^v \in \mathbb{R}^{C \times 1 \times W}$5 SRFM modules approximates global coverage in $Y^v \in \mathbb{R}^{C \times 1 \times W}$6 steps, rapidly expanding the effective field-of-view while maintaining local detail and preventing over-smoothing along the short axis (Hou et al., 2020). In (Lin et al., 17 Oct 2025), parallel global pooling branches ensure that the receptive field covers the entire horizontal or vertical axis, ideal for features such as elongated cracks with extreme aspect ratios.

3. Integration in Neural Network Architectures

SRFMs are inserted as plug-and-play modules in modern convolutional backbones:

• Semantic segmentation (Hou et al., 2020):
  • In dilated ResNet-50/101 (final stride 8), SPMs are placed after the $Y^v \in \mathbb{R}^{C \times 1 \times W}$7 conv in the last bottleneck of Res2, Res3, Res4, and after every Res5 bottleneck.
  • No change in channel count is required, facilitating seamless residual connections.
  • When combined with the Mixed Pooling Module (MPM), architectures can aggregate both global strip-based and multi-scale local context.
• Object detection (Lin et al., 17 Oct 2025):
  • SRFM replaces the C3k2 bottleneck in YOLO11-based C3k2 modules, particularly in the Neck, where P3–P5 FPN outputs pass through SRFM-enhanced blocks (termed "C3k2-SRFM").
  • In the Small-Scale Enhancement Module (SSEM), high-resolution P2 features are processed through SRFM, and strip-enhanced features are further propagated upwards and downwards in the network.

4. Computational Efficiency and Parameterization

SRFMs achieve significant computational savings compared to traditional dense kernels:

Module	Parameter Count (per block)	Primary Contrib. Operations
SPM (SRFM) (Hou et al., 2020)	$Y^v \in \mathbb{R}^{C \times 1 \times W}$8 (e.g. $Y^v \in \mathbb{R}^{C \times 1 \times W}$9, $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$0M per block)	$Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$1 1D convs (1x3, 3x1), 1x1 conv
MPM	$Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$2M – $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$3M	3x3 convs, strip pool convs, 1x1 reduce/expand
SRFM (Lin et al., 17 Oct 2025)	$Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$4, typically $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$5× fewer than $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$6 conv	DW conv, long strip convs, small strip convs, 1x1 attention

In practice, SRFM and MPM modules add $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$7 FLOPs on $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$8 input. In (Lin et al., 17 Oct 2025), replacing several standard $Y_{c,i,j} = \hat{Y}^h_{c,i,1} + \hat{Y}^v_{c,1,j}$9 convs in YOLO11 C3k2 modules with SRFM sub-blocks reduced FLOPs by 5.8% and parameter count by 5.7%, since the strip representation is more efficient for global context aggregation.

5. Empirical Results and Ablative Analysis

The impact of SRFM is validated through extensive experiments on standard and specialized benchmarks:

• Scene Parsing (ADE20K, Cityscapes, Pascal Context) (Hou et al., 2020):
  • Baseline FCN (ResNet-50, stride 8): 37.63% mIoU (ADE20K val)
  • +2 MPMs: 41.92% mIoU (+4.29)
  • Baseline + SPM only: 41.66% mIoU (+4.03)
  • Full configuration (+MPM + SPMs): 44.03% mIoU, 11.9M extra parameters
  • With ResNet-101: 45.60% mIoU
  • Cityscapes test (ResNet-101): 82.0% mIoU (SOTA at publication)
  • Pascal Context (59 classes): 54.5% mIoU (SOTA)
  • MPM ablation (ADE20K Res50): LRD (strip) only: 41.14%; combined (SRD+LRD): 41.92%
• Road Damage Detection (RDD2022) (Lin et al., 17 Oct 2025):
  • Chinese subset: Adding SRFM to YOLO11-based pipeline yields:
  • ΔF1 = +1.4 (80.2 → 81.6)
  • ΔmAP50 = +1.1 (83.3 → 84.4)
  • ΔmAP50:95 = +1.3 (50.7 → 52.0)
  • Per-category improvements: Longitudinal cracks +2.6, transverse cracks +2.1 AP50.
  • Integrating SRFM reduces parameter and GFLOP overhead, without penalizing inference speed.

Ablation studies consistently indicate that SRFM’s long-range, anisotropic context modeling is highly effective for segmenting and detecting high-aspect-ratio, elongated structures.

6. Motivation, Theoretical Rationale, and Use Cases

SRFMs are motivated by the observation that many real-world structures—such as fence-lines, roads, cracks, and urban boundaries—exhibit elongated, anisotropic geometry. Conventional square kernels inefficiently allocate parameters in both directions, failing to capture such cues without excessive cost or spatial smoothing. By focusing capacity along either the horizontal or vertical axis (via 1xK or Kx1), SRFM attains global context along the structure while preserving local detail in the orthogonal dimension.

This design directly improves network ability to discriminate, localize, and parse slender high-aspect-ratio objects and boundaries, and has been shown to outperform regular square-kernel pooling for these use cases—most notably in scene parsing and road damage detection (Hou et al., 2020, Lin et al., 17 Oct 2025).

7. Related Modules: Mixed Pooling Module (MPM)

The Mixed Pooling Module (MPM) (Hou et al., 2020) combines:

• Long-range context (via SRFM’s strip pooling branch—“LRD”)
• Medium-range context (via small bin average pooling, 20x20 and 12x12, with upsample + $X$0 conv)
• Local detail (via standard $X$1 conv)

All branches are fused (channel-wise concatenation and a $X$2 expand) to produce the output. Ablation reveals that both short- and long-range branches contribute independently to improved accuracy, with their combination consistently yielding the best segmentation metrics.

---

In summary, the Strip Receptive Field Module realizes efficient, anisotropic, and large receptive fields for modern convolutional networks, advancing dense prediction performance—especially for geometric and high-aspect-ratio targets—across a wide range of vision tasks (Hou et al., 2020, Lin et al., 17 Oct 2025).