Infrared weak-texture gas image segmentation via physics-informed learning

CH₄ gas is widely used as a clean energy source in modern industry, while its high flammability and explosiveness demand accurate leak detection. Infrared imaging provides a non-contact and robust method for visualizing gas leaks under various environmental conditions. However, CH₄ usually appears with weak textures and low contrast under low leakage conditions, making challenges for precise segmentation. We propose a segmentation learning framework informed by a dual-attention mechanism and regularized by a contrast prior. Specifically, the channel attention block enhances high-level semantic feature extraction, enabling the model to obtain weak texture details while suppressing background noise. The spatial attention block further guides the network to focus on leakage regions, improving detection of small and low-contrast targets. Additionally, the local contrast-aware regularization term derived from image gradients is integrated into the loss function to refine boundary learning and improve edge accuracy. Experimental results of an infrared CH₄ image dataset demonstrate that our method achieves an improvement of approximately 18.77% in AP50@95 compared to existing approaches, providing a potential application for precise infrared gas leak detection.