Physics-Inspired Modeling and Content Adaptive Routing in an Infrared Gas Leak Detection Network

TL;DR

This work tackles infrared gas leak detection under challenging conditions where plumes are faint and boundaries are weak. It marries physics‑inspired diffusion–convection modeling with edge‑aware, multi‑scale perception and content‑adaptive feature routing to form PEG‑DRNet, a single‑stage detector built on RT‑DETR. Key contributions include the Gas Block that emulates diffusion and convection in feature space, the AGPEO/MSEPM module for reliable edge priors, and the CASR‑PAN neck for sparse yet informative cross‑scale fusion. Experimental results on the IIG and LangGas datasets show superior AP and AP$_{50}$ with a favorable efficiency profile (e.g., $AP=29.8\%$, $AP_{50}=84.3\%$, $Gflops=43.7$, $Params=14.9\text{M}$), demonstrating practical potential for real‑time, edge‑deployable IR gas leak detection. The approach offers a physics‑guided, interpretable path to improving detection of small, semi‑transparent plumes while balancing accuracy and computational cost, with implications for safer and more efficient industrial monitoring.

Abstract

Detecting infrared gas leaks is critical for environmental monitoring and industrial safety, yet remains difficult because plumes are faint, small, semitransparent, and have weak, diffuse boundaries. We present physics-edge hybrid gas dynamic routing network (PEG-DRNet). First, we introduce the Gas Block, a diffusion-convection unit modeling gas transport: a local branch captures short-range variations, while a large-kernel branch captures long-range propagation. An edge-gated learnable fusion module balances local detail and global context, strengthening weak-contrast plume and contour cues. Second, we propose the adaptive gradient and phase edge operator (AGPEO), computing reliable edge priors from multi-directional gradients and phase-consistent responses. These are transformed by a multi-scale edge perception module (MSEPM) into hierarchical edge features that reinforce boundaries. Finally, the content-adaptive sparse routing path aggregation network (CASR-PAN), with adaptive information modulation modules for fusion and self, selectively propagates informative features across scales based on edge and content cues, improving cross-scale discriminability while reducing redundancy. Experiments on the IIG dataset show that PEG-DRNet achieves an overall AP of 29.8\%, an AP$_{50}$ of 84.3\%, and a small-object AP of 25.3\%, surpassing the RT-DETR-R18 baseline by 3.0\%, 6.5\%, and 5.3\%, respectively, while requiring only 43.7 Gflops and 14.9 M parameters. The proposed PEG-DRNet achieves superior overall performance with the best balance of accuracy and computational efficiency, outperforming existing CNN and Transformer detectors in AP and AP$_{50}$ on the IIG and LangGas dataset.

Figures (13)

• Figure 1: Comparing the AP$_{50}$ and Gflops of detection methods on IIG.
• Figure 2: The overall architecture of our proposed PEG-DRNet. In AIMM-F, the bias addition (BA) represents the learned channel-wise offsets that adaptively calibrate feature responses, while in AIMM-S, the identity-aware scaling (IDAS) denotes the identity-specific scaling factors that adjust feature amplitudes based on domain knowledge. In both modules, the information richness illustrated in the figures reflects the amount of semantic and spectral information retained in the feature maps, and the weight values are generated by the Importance Estimator, indicating the relative significance of each feature channel for downstream tasks.
• Figure 3: The Physics–Edge Hybrid Backbone integrates initial convolution and downsampling with multiple Gas Blocks. Each Gas Block is physically inspired by the convection and diffusion equation: a local branch captures diffusion-like fine-grained features, a global branch performs frequency-domain decay to model convection-like transport, and edge-aware gating modulates global features, with residual fusion producing the final output.
• Figure 4: Architecture of the multi-scale edge perception module (MSEPM) and adaptive gradient and phase edge operator (AGPEO)
• Figure 5: Architecture of the importance estimator.