Work

Gas cloud detection model based on global attention mechanism and transfer learning

Year: 2026

Type: article

Abstract: Since most hazardous gases are transparent in the visible spectrum, infrared imaging is indispensable for detection. However, in complex industrial environments, this method suffers from high false alarm rates due to irregular gas cloud boundaries. To address detection challenges and high false alarm rates caused by irregular gas cloud boundaries in complex industrial environments, this paper proposes an improved Moving Fluid Boundary YOLO (MFB-YOLO). Built upon the YOLOv8 framework, the algorithm incorporates three key enhancements to capture dynamic gas cloud characteristics. First, infrared–visible image fusion is employed to robustly differentiate transient gas clouds from stable interference sources like personnel. Second, addressing the limitations of conventional static attention mechanisms for fluid targets, we introduce a Novel Global Attention Mechanism (GAM). GAM synergistically integrates channel and spatial attention via dynamic weight allocation to prioritize gas-dominant regions, significantly improving adaptability to fluid boundaries. Third, to tackle the generalization bottleneck caused by the dynamic nature of diffusion and data scarcity, we integrate a strategy combining transfer learning with time-series analysis. This allows the model to inherit robust feature representations from large-scale data-rich domains while simultaneously capturing the continuous spatiotemporal evolution of gas shapes. Experimental results on 1,1-difluoroethane demonstrate that the improved MFB-YOLO achieves a Precision of 93% and an mAP50 of 89%, substantially outperforming the baseline. Although the inference speed increases slightly to 0.09 seconds per sample, it remains significantly faster than the image acquisition interval, ensuring real-time performance. This work presents a robust, industrially viable solution for high-precision detection of non-fixed-shape gas clouds.

Source: Optics & Laser Technology

DOI: 10.1016/j.optlastec.2026.115199

Authors: Bo Yan, Hao Chen, Fudong Li, Junyong Fang, Shuaihui Li

Institutions: Chinese Academy of Sciences, Institute of Mechanics, University of Chinese Academy of Sciences, State Key Laboratory of High Temperature Gas Dynamics, Aerospace Information Research Institute

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,230

Funder: Chinese Academy of Sciences