An Infrared Gas Imaging and Instance Segmentation Based Gas Leakage Detection Method

To realize automatic leakage detection for infrared gas imaging, we propose an instance segmentation method for gas plumes, which can simultaneously offer leakage detection, plume segmentation, and multi-source detection. To model the anisotropy of plumes in embedding space, different from the existing instance segmentation methods, we employ a new clustering loss function based on the similarity of the probability of two-dimensional Gaussians. The loss function pulls the pixels of the instance together and jointly maximizes the segmentation mask of each plume by learning a bandwidth that is with a slanted elliptical shape. Moreover, to obtain more infrared gas imaging data and avoid the difficulty of manually labeling plume contours, we generate a large synthetic infrared gas imaging data set and train our model on synthetic data. Experimental results show that our method can successfully perform automatic leakage detection on real infrared videos after training on synthetic data. Compared with the state-of-the-art methods, our method can perform plume instance segmentation at a higher speed while maintaining a high accuracy, which is suitable for real-time detection.