Multi-Scale Differential Edge Module

• The MSDE module integrates multi-scale pooling and differential operators to extract, enhance, and fuse edge signals in neural networks.
• It employs both first- and second-order operators to capture fine edge details and suppress noise across tasks like infrared detection and 3D estimation.
• Experimental results show measurable improvements in applications such as crisp edge detection and surface normal estimation from point clouds.

A Multi-Scale Differential Edge (MSDE) module is a neural network architectural component designed to extract, enhance, and fuse edge-related signals at multiple scales via differential operators. MSDE modules are widely utilized for tasks requiring precise preservation and enhancement of edge or boundary information in feature maps, such as infrared small target detection, crisp edge detection in natural images, and surface normal estimation from 3D point clouds. Distinct configurations of MSDE exist for both grid-structured and sparse-domain representations, but they share a common principle: combining multi-scale contextual feature aggregation with explicit edge emission based on the application of differential operators (first or second order).

1. Motivations and Theoretical Principles

Edge-related information in deep neural networks often degrades due to repeated pooling and convolutions, causing high-frequency details to be smoothed and faint edges eliminated. This loss is pronounced when the targets are small (e.g., a few pixels) or boundaries are weak, as in infrared small target detection (Li et al., 23 Jan 2026). Similarly, in image edge detection, standard DCNN approaches produce thick or noisy edge maps, failing to leverage edge prior knowledge (Liu et al., 2024). In 3D point cloud domains, regions with sharply varying normals correspond to "surface edges" that are obscured when only low-frequency spatial context is aggregated (Xiu et al., 2023).

To address these issues, MSDE modules:

• Implement explicit multi-scale pooling or dilation to aggregate context at several field-of-view sizes.
• Apply differential operators (classically first or second-order derivatives) to highlight local discontinuities.
• Fuse the outputs via simple summation or learned embedding schemes.
• Optionally refine with attention or adaptive gating mechanisms.

This design enables the network to retain geometrically salient features lost to downsampling, while suppressing spurious or noisy responses.

2. Canonical Architectures: Grid-Structured MSDE Modules

In grid-structured networks, such as CNNs for infrared or natural images, MSDE modules are typically inserted in the encoder, operating in parallel with the main feature extraction path (Li et al., 23 Jan 2026, Liu et al., 2024).

Example: MDAFNet MSDE Block (Li et al., 23 Jan 2026)

For a feature map $X \in \mathbb{R}^{C \times H \times W}$, the MSDE module executes:

1. Channel projection (CBS): $E_0 = \mathrm{CBS}(X)$, where CBS is $1 \times 1$ Conv + BN + Sigmoid.
2. Multi-scale edge extraction: Iteratively apply hierarchical average pooling and differential edge extraction for $t=1 \ldots T$ scales:
   • $E_t = \mathrm{AP}(\mathrm{CBS}(E_{t-1}))$
   • $$E_t^{ed} = E_t + \mathrm{CBS}(E_t - \mathrm{AP}(E_t))$$
3. Fusion: Concatenate $\{E_0, E_1^{ed}, \ldots, E_T^{ed}\}$ and apply $1 \times 1$ Conv: $E^c = U([\cdot])$.
4. Channel & spatial attention:
   • Channel: $$W_c = \sigma(\phi(\text{G-AP}(E^m)) + \phi(\text{G-MP}(E^m)))$$
   • Spatial: $$W_s = \sigma(\text{Conv}_{1\times1}(\text{SiLU}(\text{Conv}_{7\times7}([\,\text{mean},\,\text{max},\,\text{min},\,\text{sum}\,]))))$$
   • Refined: $E^{sa}_{out} = E^{ca}_{out} \odot W_s$
5. Residual multiplication: Enhanced output: $E^{out} = (E^{sa}_{out} + E^c) \odot E^c$

Example: SDMCM Block (LUS-Net) (Liu et al., 2024)

For each encoder output $X$:

1. Context path: Channel compression by $1 \times 1$ Conv; $\rightarrow$ four branches with stacked $3 \times 3$ convolutions (dilations $d=1,2,3$).
2. Second-order path: Apply $3 \times 3$ Laplacian kernel $K_L$; BN + ReLU; followed by $3 \times 3$ Conv, BN; then $1 \times 1$ Conv; residual skip from $X$.
3. Fusion: $Y = F_{context} + F_{lap}$

3. MSDE for Irregular Domains: Point Clouds

In point cloud analysis, MSDE modules operate on unordered sets, fusing multi-scale features for each point and applying an adaptive differential operator (Xiu et al., 2023).

Example: MSEC Stream (MSECNet) (Xiu et al., 2023)

1. Multi-scale fusion: Upsample features from multiple PointNet++ scales to each point; concatenate and embed; apply "space transformation" (max over $k$-NN neighborhood) and "channel transformation" (MLP); residual sum.
2. Adaptive edge detection: For each point, compute Laplacian-like difference over $k$-NN, modulated by tiny MLPs for adaptability.
   • $$e_i = \phi\left(\frac{1}{|\mathcal{N}(i)|} \sum_{j \in \mathcal{N}(i)} \theta(f_j^{ms} - f_i^{ms})\right)$$
3. Edge conditioning: Modulate backbone features by edge signals via an MLP and residual addition.
   • $f_i^{cond} = f_i^b + \gamma([f_i^b \parallel e_i])$

4. Differential Operators and Multi-Scale Contextualization

MSDE modules leverage specific mathematical operators:

• First-order: Sobel, Scharr, or discrete difference kernels. Suitable for classical edge detection but produce thick/blurred edges.
• Second-order: Laplacian operator $K_L$:

$$K_L = \begin{bmatrix} 0 & 1 & 0 \ 1 & -4 & 1 \ 0 & 1 & 0 \end{bmatrix}$$

applied as $L = K_L * X$; this achieves zero-crossing sensitivity for sharp localization (Liu et al., 2024).

• Multi-scale: Pooling, stacked dilated convolution (with rates $d=1,2,3$), and context aggregation maximize receptive fields and suppress spurious responses.

A plausible implication is that multi-scale aggregation balances sensitivity to micro-structure and robustness to noise (Xiu et al., 2023).

5. Edge Feature Fusion and Attention Mechanisms

After differential extraction, MSDE modules combine multi-scale edge signals:

• Concatenation or summation: Features extracted at different scales and differential operators are either concatenated along channels or summed. In LUS-Net, fusion is by element-wise sum, optionally with a learnable weighting (Liu et al., 2024).
• Attention refinement: Channel and spatial attention further enhance salient edge features. Deep attention mechanisms use global pooling statistics and spatial convolutions for per-channel and per-pixel weighting (Li et al., 23 Jan 2026).

6. Experimental Performance and Impact

Deploying MSDE modules in network architectures yields consistent improvements across several modalities.

Study & Task	Metric Gain with MSDE	Application Area
MDAFNet (IRSTD-1K) (Li et al., 23 Jan 2026)	IoU: +1.10%, $P_n$: +1.02%, $F_a$: –3.04e–6	Infrared Small Target Det.
LUS-Net (BSDS500) (Liu et al., 2024)	C-Eval ODS: +1.2 pts (0.698→0.686)	Crisp Edge Detection
MSECNet (PCPNet) (Xiu et al., 2023)	Angle-RMSE: –0.74° (w/ MSEC)	3D Normal Estimation

MSDE modules sharpen edges, prevent boundary attenuation, and suppress both low-frequency background and high-frequency noise. Increased numbers of scales improve performance up to a point, after which feature redundancy reduces benefits. In ablations, second-order derivatives outperform first-order for boundary localization; fusing multi-scale context mitigates Laplacian-induced noise (Liu et al., 2024).

7. Module Variants, Implementation, and Prospective Extensions

Key implementation choices include:

• Number of scales: Empirically, three to four scales suffice before redundancy arises.
• Fusion method: Element-wise sum yields robust results; learnable weights introduce flexibility.
• Attention schemes: Channel and spatial attention via global pooling/statistics refines outputs.
• Residual connections: Skip links in context and Laplacian paths reinforce feature integrity.

MSDE modules are, by construction, lightweight, plug-and-play, and provide measurable gains without explicit post-processing. While current MSDE variants operate at fixed scales and with static fusion weights, a plausible extension is adaptive learned scale weighting via attention. Additionally, formal adaptation to non-grid domains (e.g., multi-scale graph Laplacians) is possible but not present in the surveyed work (Xiu et al., 2023).