PEG-DRNet: Hybrid Gas Dynamics Routing

• The paper introduces PEG-DRNet, a neural architecture that fuses physics-inspired gas modeling with robust multi-scale edge extraction for effective infrared gas leak detection.
• It employs a novel CASR-PAN module that adaptively aggregates multi-scale features based on spatial content and edge cues, reducing computation by over 50% compared to conventional designs.
• Empirical results show enhanced average precision and small-object detection on infrared leak benchmarks, underscoring its potential in real-time safety monitoring.

The physics-edge hybrid gas dynamic routing network (PEG-DRNet) is a neural network architecture designed for infrared gas leak detection, with innovations motivated by gas transport physics, multi-scale edge perception, and content-adaptive cross-scale information routing. PEG-DRNet integrates physical modeling of gas diffusion and convection, robust edge extraction under weak-contrast conditions, and a novel content-driven, sparsely gated feature aggregation "neck" to enhance detection accuracy and efficiency, particularly for small, faint, and diffuse gas plumes (Li et al., 29 Dec 2025).

1. Key Architectural Components and Pipeline

PEG-DRNet comprises three principal modules: (1) a Gas Block for physics-inspired feature extraction, (2) an edge perception pipeline with the adaptive gradient and phase edge operator (AGPEO) feeding into a multi-scale edge perception module (MSEPM), and (3) the content-adaptive sparse routing path aggregation network (CASR-PAN).

1. Gas Block: Models gas plume evolution using a diffusion-convection scheme. The block incorporates a local branch for short-range feature dynamics and a large-kernel branch for long-range transport. A learnable, edge-gated fusion module unifies these responses, improving weak-plume and contour cues.
2. Edge Perception: AGPEO estimates reliable edge priors by combining multi-directional gradients and enforcing phase-consistency across features. Derived edge maps are passed through MSEPM, which constructs a hierarchical representation of edge cues crucial for reinforcing ambiguous boundaries.
3. CASR-PAN: The "neck" replaces rigid, predetermined fusions (as in FPN/PANet) with a content-driven, scale-wise adaptive, and sparsely gated fusion mechanism. CASR-PAN fuses multi-scale features from the backbone according to spatially varying content and edge cues, with explicit fusion paths: deep-to-mid, deep-to-shallow, shallow-to-mid, and self-enhancement.

2. CASR-PAN: Content–Adaptive Sparse Routing Path Aggregation Network

CASR-PAN accepts multi-scale feature maps $\{F_2, F_3, F_4, F_5\}$ and dynamically learns the regions and paths for information propagation based on feature importance, reducing redundancy and increasing discriminability across scales. The core workflow encompasses:

• Importance Estimator (IE): Computes three complementary cues per spatial location and channel:
  • Global ($\overline{G}$): Aggregated via global average pooling, processed by 1×1 convolutions and nonlinearity.
  • Local ($L$): Extracted with a 3×3 convolution, nonlinearity, and upsampling.
  • Diversity ($D$): Based on channel-wise standard deviation, normalized via a learnable transformation.
  • The final importance map $I \in [0,1]^{B \times C \times H \times W}$ merges these cues using softmax-weighted summation and a sigmoid function.
• Routing Weights ($W$): Four explicit paths are established, each with a spatially varying mask $W_k$ (for $k=1$ to $4$), derived via a 1×1 convolution over $I$ and sigmoid activation.
• Fusion Modules (AIMM-F, AIMM-S): AIMM-F mixes two features $\overline{G}$0, $\overline{G}$1 using a gated combination:

$\overline{G}$2

where BA is a fixed bias (0.5), and $\overline{G}$3 is the sigmoid. AIMM-S is used for self-enhancement:

$\overline{G}$4

with $\overline{G}$5.

• Aggregation and Refinement: The fused features per path are summed and passed through a RepC3 residual block to stabilize learning and further refine the representation. The resulting feature maps are consumed by a DETR-style decoder head for object detection.

3. Routing Formulation and Sparsity Induction

The routing mechanism implements, for each spatial location $\overline{G}$6 and path, a convex combination of transported and local features. For the deep-to-mid path, this is

$\overline{G}$7

Stacking all four routing weights yields a sparse block routing tensor $\overline{G}$8. Sparsity is achieved implicitly as follows:

• The use of sigmoid activations encourages each $\overline{G}$9 to saturate near 0 or 1 for many spatial locations.
• Bias addition BA ensures no path collapses entirely.
• Optionally, explicit sparsity penalties (e.g., $L$0) or post-hoc thresholding may be applied, but were not used in the main implementation.

The result is an efficient, adaptive routing graph that prunes unnecessary cross-scale communication during inference, targeting only regions or scales with salient content or edge cues.

4. Pseudocode and Data Flow

The following summarizes the CASR-PAN forward pass logic as described in (Li et al., 29 Dec 2025):

$L$4

This mechanism ensures that only the most relevant spatial regions and channels receive synthesized multi-scale support.

5. Computational Complexity and Efficiency

CASR-PAN significantly reduces computational cost in comparison to prior "neck" designs such as PANet, BiFPN, and NAS-FPN. For IIG dataset experiments, CASR-PAN requires 45.8 Gflops and 16.94 million parameters, yielding:

• A ∼56% reduction in Gflops relative to PANet (103.8 Gflops, 21.46M params)
• Lower Gflops and parameters than BiFPN (64.3 Gflops, 20.3M params) and NAS-FPN (93.8 Gflops, 19.76M params)
• Per-pixel IE and gating costs are minor compared to the overall backbone.

The overall PEG-DRNet model (including backbone and neck) achieves 43.7 Gflops and 14.9M parameters.

6. Empirical Performance and Ablations

PEG-DRNet achieves superior detection quality on challenging infrared leak benchmarks:

• On the IIG dataset: AP = 29.8%, AP$L$1 = 84.3%, small-object AP = 25.3%
• Compared to RT-DETR-R18 baseline: +3.0% AP, +6.5% AP$L$2, +5.3% small-AP
• Outperforms other CNN and Transformer detectors on IIG and LangGas

Ablation studies for CASR-PAN reveal each routing path's contribution:

• Removing deep→mid reduces AP to 27.5%
• Removing deep→shallow or shallow→mid reduces AP to 25.7%
• Removing mid-self reduces AP to 28.3%
• Full CASR-PAN shows optimal results (AP = 29.4%, AP$L$3 = 12.3%)

Compared to other neck designs:

• PANet: AP = 27.3%
• BiFPN: AP = 27.0%
• NAS-FPN: AP = 25.1% CASR-PAN uniquely achieves higher AP at substantially lower computational cost.

7. Technical Significance and Outlook

PEG-DRNet establishes a new paradigm in detector design for tasks exhibiting weak boundaries, small object size, and strong multi-scale dependency. Its integration of physics-inspired gas modeling, robust edge-guidance, and data-driven, sparse multi-scale fusion leverages domain priors and adaptive learning. The CASR-PAN neck demonstrates that content-aware, sparsity-inducing routing can yield both computational and accuracy gains, making the approach especially suitable for applications in environmental monitoring and industrial safety where real-time, high-fidelity plume detection is vital (Li et al., 29 Dec 2025).