DAEiS-Net: Deep Aggregation Network with Edge Information Supplement for Tunnel Water Stain Segmentation

Area of Science:

• Urban engineering
• Civil engineering
• Computer vision

Background:

• Tunnel disease detection and maintenance are vital for urban transportation safety.
• Water stain detection in tunnels is challenging due to variable morphology, scale, and loss of contextual and boundary information.
• Existing methods struggle with multiscale feature extraction and edge detail in complex tunnel environments.

Purpose of the Study:

• To propose a novel deep learning method, DAEiS-Net, for accurate tunnel water stain detection.
• To enhance the extraction of multiscale contextual and edge information for improved segmentation accuracy.
• To introduce a new dataset, TWS, for training and evaluating tunnel water stain segmentation models.

Main Methods:

• The study introduces the Deep Aggregation Network with Edge Information Supplement (DAEiS-Net), utilizing an encoder-decoder architecture.
• Key modules include Deep Aggregation Module (DAM) for feature enhancement, Multiscale Cross-Attention Module (MCAM) for noise suppression and texture enhancement, and Edge Information Supplement Module (EISM) to bridge semantic gaps.
• A Sub-Pixel Module (SPM) is employed for multiscale feature fusion and edge representation enhancement.

Main Results:

• DAEiS-Net demonstrated superior performance in tunnel water stain segmentation on the newly introduced Tunnel Water Stain Dataset (TWS).
• The proposed modules effectively addressed challenges related to multiscale information extraction and boundary detail.
• Experimental results confirm the state-of-the-art capabilities of DAEiS-Net for this specific task.

Conclusions:

• DAEiS-Net offers a significant advancement in automated tunnel water stain detection and segmentation.
• The method's ability to integrate multiscale features and edge information leads to more robust and accurate results.
• The TWS dataset provides a valuable resource for future research in tunnel defect detection.