Optimised 3D-CNN for Real-Time Infrared Natural Gas Leak Classification: Balancing Accuracy and Computational Cost

Infrared imaging is vital for real-time methane leak detection in industrial environments, yet the computational complexity of 3D Convolutional Neural Networks (3D-CNNs) like VideoGasNet limits their deployment on resource-constrained systems. This paper proposes an optimised preprocessing pipeline for VideoGasNet, replacing the computationally intensive moving median background subtraction with a running average approach and incorporating image downscaling. Experiments on the GasVid dataset demonstrate that our method reduces preprocessing time significantly—by up to 66% at reduced resolutions—while maintaining high classification accuracy (above 98% even at quarter resolution). These enhancements enable real-time gas leak classification, offering a practical balance between accuracy and computational cost for industrial applications.