A Gas Detection Method Based on Multiscale Infrared Image Semantic Segmentation

Infrared imaging systems have been widely applied in gas leak detection. However, The existing gas detection methods have many limitations and are difficult to apply in real-world scenarios. At the same time, there are very few methods that combine gas detection and semantic segmentation with deep learning. In this study, a novel approach for gas detection using image semantic segmentation in deep learning is proposed. This method presents a new multi-scale semantic segmentation model named PUNet, based on PSPNet and U-Net, for automatic segmentation of infrared gas leakage images. Meanwhile, to solve the problems of single scene and fixed leakage location in the gas leakage image dataset, we added more self-collected infrared gas leakage images to the existing dataset. The experimental findings demonstrate that PUNet has higher accuracy than traditional foreground segmentation algorithm and outperforms the conventional U-Net model in segmenting gas leakage images, and exhibits enhanced efficacy in handling multi-scale gas leakage scenarios.