YOLOGAS: An Intelligent Gas Leakage Source Detection Method Based on Optical Gas Imaging

Abstract

Industrial gas leakage measurement is a guarantee for safe production. Optical gas imaging is widely used due to its advantages of wide detection range and noncontact, but it has problems, such as low detection efficiency and inability to detect intelligently. Therefore, this article proposes an automatic gas leakage detection model, named YOLOGAS. First, Swin-Transformer is introduced to achieve the effective use of global context information, so that the network can better perceive gases. Second, the convolutional block attention module (CBAM) and the efficient channel attention (ECA) module enable the network to focus on gas targets. Finally, the bidirectional feature pyramid network (BiFPN) is used to reduce the computational complexity. Additionally, based on the similarity between gases and smoke, this article uses Pix2pix network to generate infrared images for model pretraining. The experimental results show that YOLOGAS has an AP50 of 88.07% and an FPS of 21, which is superior to other detection models.