Contrast gas detection: Improving infrared gas semantic segmentation with static background

Accurate and efficient identification and segmentation of gases play an important role in industrial processes and public safety. Although Optical Gas Imaging (OGI) has proven effective in acquiring images and videos of gas leaks, accurately determining the leakage points and their diffusion areas in real-world conditions remains a significant challenge. To address this, we propose a novel gas segmentation annotation methodology. This approach applies differencing between gas leakage scenes and corresponding static background to delineate precise leakage regions, which are subsequently annotated manually. The resulting dataset, consisting of 926 static background-gas image pairs extracted from 38 comprehensive video sequences, provides a robust foundation for advancing gas segmentation research. Based on this dataset, we present a novel semantic segmentation model specifically designed for the unique characteristics of gas leakage scenarios. Our model integrates a gas contrast attention mechanism to capitalize on static background information, resulting in improved segmentation precision. Comparative evaluations demonstrate that the proposed model achieves state-of-the-art performance on our dataset, outperforming widely adopted semantic segmentation models. All code and the dataset will be made publicly available at https://github.com/ProAlize/CGNet.