Visual detection method of tunnel water leakage diseases based on feature enhancement learning

Detecting water leakage is vital for assessing tunnel structure operational conditions. Currently, deep learning (DL) based methods for leakage detection have shown promising results. However, their robustness in complex backgrounds remains limited due to challenges in extracting essential features from leakage areas. To tackle this issue, a novel detection model for water leakage is proposed, based upon feature enhancement learning. The model adopts Mask R-CNN as its core framework and seeks to enhance detection performance through three strategies as follows. Firstly, using the brightness aggregation of leakage pixels, Otsu method is initially used to segment leakage pixels. The segmented outcome is employed alongside the original image for network input, which can offer guided training to the recognition network and enhance its ability to separate leakage from backgrounds effectively. Secondly, considering the perception difference across feature extraction layers in DL networks, Non-Local Block is integrated into low-level networks, correlating leakage areas and global pixels. Additionally, Squeeze-and-Excitation Block is proposed to amplify channel weights for leakage in high-level networks, augmenting its ability to perceive crucial characteristics within leakage regions. Thirdly, addressing the issue of insufficient leakage boundary feature perception by unidirectional pyramids in existing networks, we present a Bidirectional Feature Pyramid Network. Besides, this proposed model applies one distinct inter-layer feature fusion based on the pyramid's direction. The algorithm's performance is evaluated using a tunnel leakage dataset. Through conducting ablation experiments, it was verified that the proposed model consistently outperforms other comparison algorithms in leakage detection accuracy.