A Deep Learning-Based Multi-Angle Gas Leak Detection Method

Hazardous chemicals play a crucial role in advancing industrial development; however, during storage, the potential risk of leaks presents a serious threat to both human health and the environment. To ensure effective management and monitoring of hazardous chemical storage, this paper proposes a gas leak detection method tailored for dynamic viewpoints. A Dynamic Foreground Extraction Module (DFEM) is designed to efficiently extract gas foreground images. Additionally, the Leakage Gas Detection Network (LGDNet) is optimized: initially, raw and foreground images are input in parallel to integrate the strengths of different layers, enhancing feature extraction. Furthermore, multi-scale feature fusion and an attention mechanism are incorporated to enable the model to adapt to leak scenarios at varying scales and focus on critical regions, ensuring accurate gas detection. Experimental results show that the LGDNet model with two blocks performs optimally. On the custom dataset, the LGDNet model achieved an accuracy of $99.42 \%$ and an F1 score of $99.31 \%$, offering a better balance between lightweight design and performance, surpassing mainstream deep convolutional networks. Additionally, tests on the GasVid dataset demonstrate that its accuracy exceeds that of the GasNet model by 2 percentage points, showcasing robust detection performance and adaptability.