MBA-Net: Masked Background Alignment for Robust RGB-Thermal Invisible Gas Segmentation

Abstract

Industrial safety monitoring requires reliable detection of fugitive gas emissions, which are typically invisible in RGB images but observable in thermal infrared imagery. RGB–thermal (RGB-T) fusion therefore provides a promising solution for area-based gas leak detection. However, existing CNN-based approaches are limited by local receptive fields, making it difficult to capture the diffuse and irregular structures of gas plumes. Although recent Transformer-based methods improve global context modeling, they often amplify background thermal artifacts, leading to high false-positive rates in complex industrial environments. To address these challenges, we propose MBA-Net, a dual-stream Transformer-based framework for robust RGB-T invisible gas segmentation. MBA-Net first employs a dual-stream backbone to extract multi-scale contextual features from RGB and thermal modalities. A Thermal Artifact Suppression Gate (TASG) is introduced to perform structure-guided suppression of thermally salient but structurally inconsistent background responses. To further reduce residual background bias, we design a Masked Background Alignment (MBA) loss that enforces cross-modal feature consistency in background regions during training, without introducing additional inference cost. Finally, a Confidence-Aware Refinement (CAR) module is proposed to adaptively enhance uncertain regions, improving the representation of diffuse gas boundaries and weak plume responses. Extensive experiments on the public Gas-DB benchmark demonstrate that MBA-Net achieves superior segmentation performance with competitive computational efficiency.