Gas Leakage Segmentation in Industrial Plants

Industrial gas leakage in chemical plants presents a high risk to health and safety, also makes a huge waste of production materials. In this paper, we propose an automatic gas segmentation method for optical gas imaging (OGI) videos, which can not only indicate the existence of leakage in frames but also show the precise locations and shapes of gas plumes. Specifically, we propose a novel video segmentation network that stacks 2D spatial convolutions, 1D temporal convolutions and 3D spatial-temporal convolutions together within an encoder-decoder structure, termed 2.5D-Unet. Stacking mixed convolutions increases the representation ability of network for leakage's appearance and motion. More importantly, stacking mixed convolutions facilitates pre-training of model using still images of smoke, which is especially useful when lacking of labeled videos of leaked gas. Experimental results show, for OGI video-based gas leakage segmentation, our method outperforms existing video segmentation methods.