Multi-Scale Edge Perception Module

• MSEPM is an architectural component that extracts, enhances, and fuses edge features at multiple spatial resolutions to improve boundary detection.
• It employs multi-scale pooling, parallel dilation, and channel attention mechanisms to robustly capture structural nuances in diverse vision tasks.
• Empirical studies report performance gains such as improved AP and reduced RMSE in applications from gas leak detection to retinal vessel segmentation.

A Multi-Scale Edge Perception Module (MSEPM) is an architectural component designed to extract, enhance, and propagate edge-related cues at multiple spatial resolutions within a neural network backbone. MSEPMs are deployed in diverse vision tasks—including segmentation, detection, and surface normal estimation—where capturing boundaries at multiple scales is essential for robust performance in the presence of noisy or ambiguous regions, variable object sizes, and fine structural transitions.

1. Foundational Principles and Motivations

Edge cues represent localized spatial changes in signals or features, often corresponding to object boundaries or surface discontinuities. In challenging settings such as diffuse boundaries (e.g., atmospheric plumes), low-contrast retinal vasculature, or regions of rapid normal variation in 3D geometry, single-scale or fixed-receptive field edge detectors are prone to missing important transitions or amplifying noise. The MSEPM addresses this by aggregating edge evidence across multiple spatial resolutions and contexts, fusing finer and coarser edge features to improve boundary localization and semantic disambiguation (Li et al., 29 Dec 2025, Ouyang et al., 18 Apr 2025, Xiu et al., 2023).

2. Core Architectural Variants

Multiple MSEPM instantiations have emerged, each adapted to domain requirements and backbone architectures. Three dominant forms have been proposed:

• MSEC Stream in Normal Estimation: In MSECNet, the Multi-Scale Edge Conditioning (MSEC) stream is introduced parallel to a PointNet++-style decoder. Intermediate features at multiple upsampled scales are fused and processed to yield a spatially and channel-wise aggregated representation, which is then used for adaptive edge detection and subsequent conditioning of the backbone feature via a residual module (Xiu et al., 2023).
• MSEPM in Gas Leak Detection: Within PEG-DRNet, the MSEPM follows the Adaptive Gradient and Phase Edge Operator (AGPEO), hierarchically pooling a fused edge prior to create channel-mixed edge feature maps at four progressively coarser resolutions. These are injected into the subsequent backbone stages via attention-based summation (Li et al., 29 Dec 2025).
• Retinal Edge Fusion Enhancement: The MREF module in high-resolution vessel segmentation frameworks applies multi-dilated convolutions in parallel, aggregates the resulting edge features by summation, and further enhances them with channel-wise Squeeze-and-Excitation (SE) attention. This approach can be generalized as an MSEPM architecture for edge enhancement in dense prediction tasks (Ouyang et al., 18 Apr 2025).

3. Multi-Scale Edge Extraction and Fusion Strategies

A canonical MSEPM integrates feature processing at several stages:

• Multi-Scale Feature Generation:
  • Pooling-Based: Spatial max-pooling recursively reduces resolution, generating a feature pyramid for different scales ((Li et al., 29 Dec 2025); equations for $E_i$ and $\hat E_i$).
  • Parallel Dilation: Multiple parallel branches with different dilation rates capture edges corresponding to small and large structures (Ouyang et al., 18 Apr 2025).
  • Multi-level Backbone Taps: Features at distinct depths or after upsampling are separately processed and then unified (Xiu et al., 2023).
• Channel Mixing and Attention:
  • $1\times1$ convolutions are commonly applied after scale-wise pooling or convolution to align channel ranks and facilitate fusion (Li et al., 29 Dec 2025, Ouyang et al., 18 Apr 2025).
  • Channel-wise attention mechanisms, such as the SE block, allow the model to suppress noise and emphasize informative edge features (Ouyang et al., 18 Apr 2025).
• Residual and Conditional Integration:
  • Edge-conditioned features are often added back to, or concatenated with, the backbone features in a residual manner to refine representations without losing baseline predictions (Xiu et al., 2023).

4. Mathematical Formalization and Algorithmic Flows

Specific algorithmic flows and formulations are given below, isolating the universal template while respecting domain variants.

a) Multi-Scale Pooling and Channel Mixing

From (Li et al., 29 Dec 2025), given an initial edge map $E_0$: $$E_i = \begin{cases} E_0, & i=0 \ \mathrm{MaxPool}_{2\times2}(E_{i-1}), & i\ge 1 \end{cases}, \quad \hat E_i = \mathrm{Conv1\times1}(E_i)$$ for multi-scale edge features at scale $i$.

b) Parallel-Dilation Fusion with SE Attention

From (Ouyang et al., 18 Apr 2025): $$\begin{aligned} \mathbf{D}_j &= \mathrm{ELU}(\mathrm{Conv}(\mathbf{Z}), \text{dilation}=d_j) \ \mathbf{F}_\text{edge} &= \sum_{j=1}^N \mathbf{D}_j \ \mathbf{z} &= \text{GlobalAvgPool}(\mathbf{F}_\text{edge}) \ \mathbf{s} &= \sigma(W_2\,\mathrm{ReLU}(W_1\,\mathbf{z})) \ \mathbf{F}_\text{enhanced} &= \mathbf{F}_\text{edge} \odot \mathbf{s} \ \mathbf{Q} &= \mathrm{Conv}_{1\times1}(\mathbf{F}_\text{enhanced}) \end{aligned}$$ where $N$ is the number of dilation rates.

c) Adaptive Edge Detection (MSEC Stream)

From (Xiu et al., 2023): $$e_i = \varphi\Bigl(\frac1{|\mathcal{N}(i)|} \sum_{j\in\mathcal{N}(i)} \theta(f^{ms}_j - f^{ms}_i)\Bigr)$$ enables adaptive, noise-robust spatial edge response computation.

5. Integration and Training Paradigms

MSEPMs are typically integrated with the backbone at multiple junctions, enabling both shallow and deep convolutional layers to receive edge cues. For example:

• In PEG-DRNet, multi-scale edge features $\{\hat E_i\}$ are fused with backbone outputs $\hat E_i$0 in the AIFI module using attention mechanisms (Li et al., 29 Dec 2025).
• In MSECNet, the edge-conditioned features are linearly combined with backbone features before final regression (Xiu et al., 2023).
• In vessel segmentation networks, MSEPMs enhance each stage's output prior to prediction, following the high-resolution backbone philosophy (Ouyang et al., 18 Apr 2025).

No edge-specific auxiliary loss is used; training proceeds end-to-end with task-appropriate supervision (e.g., standard detection loss, normal regression, segmentation cross-entropy) (Li et al., 29 Dec 2025, Xiu et al., 2023, Ouyang et al., 18 Apr 2025).

6. Empirical Impact and Ablation Studies

Ablation studies across domains demonstrate the universal utility of MSEPMs:

Configuration	Gain in Main Metric	Study / Task
+AGPEO/MSEPM vs Baseline	+1.1 AP, +4.1 AP$\hat E_i$1	Gas Leak Detection (Li et al., 29 Dec 2025)
+MSEC Stream vs Backbone	RMSE $\hat E_i$20.74°	Surface Normal Estimation (Xiu et al., 2023)
+MREF vs HR Baseline	Dice +0.2%, clDice +0.1%	Retinal Vessel Segmentation (Ouyang et al., 18 Apr 2025)

MSEPMs provide the largest improvements in low-contrast, weak-boundary, or high-frequency edge regions. For instance, in (Xiu et al., 2023), error heatmaps confirm the greatest advantage accrues at sharp edges and corners in point clouds. Similarly, (Li et al., 29 Dec 2025) finds MSEPM particularly effective for irregular, faint plumes.

Component-wise ablation underscores the necessity of multi-scale fusion, spatial/channel transformations, and adaptive attention. Omitting any branch or scale, removing the attention mechanism, or bypassing edge conditioning typically produces measurable performance degradations.

7. Applications and Generalization

MSEPMs have demonstrated broad generalization potential:

• 3D Surface Normal Estimation: Enhances detection of surface discontinuities, leading to state-of-the-art RMSE reductions (Xiu et al., 2023).
• Infrared Gas Leak Detection: Reinforces faint and low-SNR boundaries, yielding superior AP and small-object AP (Li et al., 29 Dec 2025).
• Medical Image Segmentation: Improves thin vessel tracking, maintains structural continuity, and suppresses noise in retinal images (Ouyang et al., 18 Apr 2025).

For cross-domain adaptation, the MSEPM recipe involves branching feature extraction by receptive field or pooling depth, summing or weighted-fusing edge feature branches, applying channel and/or spatial attention, and using residual integration with the backbone features (Ouyang et al., 18 Apr 2025). Design parameters—such as choice of dilation rates, branch weighting method, and scale balancing—are adjusted to match the characteristic scales and boundary clarity of the target domain.

In summary, the Multi-Scale Edge Perception Module represents a class of architectures designed to propagate, enhance, and fuse edge cues at multiple spatial scales, enabling improved structural discrimination in numerous vision tasks, especially where precise boundary information is crucial (Li et al., 29 Dec 2025, Xiu et al., 2023, Ouyang et al., 18 Apr 2025).