An improved leakage detection method for underground engineering based on dual-spectral image fusion with enhanced salient features

Computer vision methods based on multi-source data fusion are increasingly employed to detect apparent defects in underground engineering, such as water leakage. However, traditional image fusion-based detection methods often suffer from limitations such as loss of details or introduction of noise. To overcome these drawbacks, this study proposes an enhanced leakage detection method utilizing dual-spectral image fusion with saliency feature enhancement. Initially, a dataset with 8,856 pairs of visible-infrared leakage images, accounting for complex environmental factors, is constructed using data augmentation techniques. Subsequently, the YOLOv8 model is enhanced by incorporating an attention mechanism module and a feature extraction module. Finally, a dual-spectral image fusion method based on saliency feature enhancement is proposed, integrating the features of infrared and visible light images. The results demonstrate that the integration of GAM attention mechanisms and ASPP feature extraction modules significantly improves the YOLOv8 model's performance. Additionally, compared to the YOLOv8 model trained with visible light images, the model with the proposed image fusion method shows notable improvements: accuracy increases by 0.022, recall by 0.015, F1 score by 0.018, and mAP@0.5:0.95 value by 0.093. Furthermore, the proposed image fusion method proves to be effective in identifying leakage by being applied to Faster R-CNN and Retina Net. This study provides a valuable method to improve the leakage detection accuracy for underground engineering.