Lightweight Single-Stage Network for Gas Leak Detection Based on Infrared Imaging

Gas leak detection is essential for real-time monitoring and safety early warning of industrial production, manufacturing and transportation processes. For many years, infrared optical gas imaging has been widely used in the field of gas leak monitoring, but the task still faces great challenges due to the limitations of infrared imaging principle and system technology, as well as the characteristics of insubstantial gas objects. First, a dataset containing 66,950 infrared images is built, which covers gas leak samples with different scales, shapes and blurring levels. Second, a single-stage gas leak detection network model named Dual Layer Focus Aggregation Network (DLFANet) was designed. Specifically, a lightweight feature extraction cross-stage partially efficient two-layer aggregation network (CSP-EDLAN) module is designed to enhance the transmission of gradient flow information and cross-channel information interaction, where dual convolution (DualConv) is utilized to reduce the computational consumption of feature extraction. A focal modulation module is introduced into the backbone network to realize the focus of the gas target by integrating the characteristic information of different scales. In addition, The Wise Intersection Shape Intersection over Union (Wise-Shape-IoU) loss function with a dynamic non-monotonic mechanism and shape constraint capability is designed to prevent low-quality samples from generating harmful gradients, which makes the bounding box regression (BBR) of gas targets with greater accuracy. Finally, extensive experimental results on the constructed dataset show that the proposed DLFANet strikes a better balance between detection accuracy (map) and speed (FPS) while predicting the BBR of gaseous objects more accurately compared to state-of-the-art models.