Class-imbalanced gas thermal image detection in manufacturing using thermal region masking and density-based reweighting

Abstract

Recently, research on data-driven gas detection has been actively conducted to ensure the safety of manufacturing processes efficiently. Still, there remain challenges in real-world gas detection applications due to a lack of labeled data and class imbalance. While a few works have addressed these challenges through sampling or cost-sensitive methods, the characteristics of gas thermal image data, such as high noise and low resolution, hinder the application of existing data-driven methods. Therefore, this work proposes a practical gas detection method suitable for thermal image data to overcome existing limitations. First, contrastive learning (CL) is applied to learn informative latent representations using unlabeled data, which are easily collected in automated manufacturing execution systems. Second, a novel data augmentation module suitable for CL with thermal images is proposed, which generates effective views by removing uninformative backgrounds. Third, a reweighting technique is proposed to address class imbalance, utilizing a density estimation module that leverages latent representations obtained through CL. The effectiveness of the proposed method is validated on both the gas thermal image and the GasVid datasets. Compared to baseline ResNet models, the proposed method improves accuracy by 7.1 % and 6.4 % on the respective datasets. In addition, the proposed method outperforms SimCLR by 3.5 % and 3.3 %. The proposed method's ability to operate without extensive labeling and adaptability to variable input conditions makes it suitable for real-world deployment. In particular, the proposed method shows potential for scalable integration into industrial systems, enabling timely detection of gas leaks and equipment damage, thereby ensuring workers' safety.