Flow Faster RCNN : Deep Learning Approach for Infrared Gas Leak Detection in Complex Chemical Plant Surroundings

Leakage of hazardous gases from chemical plants brings serious risks to people's lives and property, therefore the infrared gas leakage detection method has attracted more and more researchers' attention with its real-time and accurate detection performance. However, infrared gas leak databases are very difficult to collect due to the flammability, explosibility and toxicity of leaked gas. Given this, we propose a gas leak detection method for chemical plant environment, which starts from both data enhancement and model improvement. Firstly, we generate and manually label fake infrared gas leak datasets under complex scenarios(ComplexGasVid dataset) as model training datasets. Secondly, we investigate the impact of the generated ComplexGasVid dataset on the performance of the target detection model. Meanwhile, we propose a new gas leak detection that combines the motion information of optical flow images with a Faster RCNN network (named Flow Faster RCNN). The final experimental results show that our proposed Flow Faster RCNN network improves the detection performance by 15.6 % compared to that of the original network.