A cross-stage adaptive fusion network for gas leak detection based on infrared imaging

Abstract To mitigate safety risks associated with gas leaks, the rapid and accurate identification of gas leakage sources in infrared imagery has become increasingly crucial. In this study, the single-stage detection model YOLOv8 was employed, as it effectively predicts target centers and demonstrates strong adaptability in complex monitoring environments. An innovative ConvFormer module was introduced, integrating depthwise separable convolutions with a MLP (Multilayer Perceptron) architecture to efficiently extract both local texture and global structural features of gas plumes. Furthermore, a CAAFM (Cross-stage Adaptive Attention Fusion Module) was developed, which employs a cross-stage adaptive attention fusion mechanism to suppress background interference and enhance the representation of small-scale targets. Additionally, a WGRM (Wavelet Group Reconstruction Module) was designed as an upsampling component to reconstruct weak edges and preserve high-frequency details in infrared gas imagery. To evaluate the effectiveness of the proposed method, experiments were conducted on an infrared gas image dataset. Compared with state-of-the-art approaches, the proposed model achieved a mAP of 47.6%, a precision of 68.6%, and a recall of 42.3%. These results confirm that the model attains a favourable balance between computational efficiency and detection accuracy.