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

Yuliang Wang, Kai Huang, Kai Zheng, Shuliang LiuSensors 2024 · OpenReview

TL;DR

This paper introduces DAEiS-Net, a novel deep learning architecture for segmenting water stains in tunnel images. By incorporating specialized modules for aggregating multi-scale features and enhancing edge information, DAEiS-Net achieves state-of-the-art performance on a newly created dataset, addressing critical challenges in tunnel infrastructure monitoring.

What Problem Does This Paper Address?

Maintaining the safety and stability of urban transportation systems relies heavily on the timely detection and repair of tunnel diseases. Water stains, a common issue, pose a significant challenge for automated detection due to their highly variable shapes, sizes, and appearances, often exacerbated by poor lighting and noise within tunnels. Existing methods struggle to effectively capture both the broad context of these stains and their precise boundaries, leading to inaccurate segmentation. This paper aims to bridge this gap by developing a more robust and accurate method for tunnel water stain segmentation.

Methods

The proposed DAEiS-Net builds upon a standard encoder-decoder architecture, similar to U-Net, but introduces several novel modules to enhance its capabilities:

  • Deep Aggregation Module (DAM): Integrated into the encoder, this module aims to improve the network's ability to represent features by fusing information from different stages.
  • Multiscale Cross-Attention Module (MCAM): This module is designed to suppress noise in shallow features and enhance texture information in deeper features by adaptively attending to multi-scale contextual information. It combines spatial and semantic attention mechanisms.
  • Edge Information Supplement Module (EISM): This module specifically targets the problem of losing boundary details. It bridges semantic gaps between different feature extraction stages by combining high-level semantic information with low-level regional texture information to better delineate the edges of water stains.
  • Sub-Pixel Module (SPM): Used in the decoder, this module enhances the representation of edge features by effectively fusing features at various scales, aiming to minimize information loss during upsampling, a common issue with traditional methods like bilinear interpolation.

Datasets

The authors introduce a new dataset called the Tunnel Water Stain Dataset (TWS), specifically curated for this task. It comprises 2000 training images, 400 validation images, and 100 testing images, capturing a diverse range of tunnel water leakage scenarios and conditions. The images have resolutions of 2330 × 1747 and 1944 × 2592, and each image is meticulously annotated with water stain contours.

Key Insights and Ideas

  • The importance of multi-scale feature fusion: Water stains vary greatly in size and shape, making it crucial for a model to effectively capture information across different scales. The DAM and MCAM modules directly address this by aggregating and attending to multi-scale features.
  • Edge detail is critical for accurate segmentation: Simply identifying the presence of a stain isn't enough; precise boundary delineation is essential. The EISM module's focus on bridging semantic gaps to enhance edge information is a key innovation.
  • Sub-pixel operations can improve decoder performance: Traditional upsampling methods can lead to information loss and blurring. The SPM module's use of sub-pixel operations offers a more detail-preserving way to reconstruct feature maps in the decoder.
  • A dedicated dataset is crucial for specialized tasks: The creation of the TWS dataset highlights the need for domain-specific benchmarks to properly evaluate and advance research in niche areas like tunnel disease detection.

Key Results

DAEiS-Net achieved state-of-the-art performance on the TWS dataset. Quantitatively, it reached a mean accuracy (mAcc) of 0.7923, a mean Intersection over Union (mIoU) of 0.7874, and a mean F1-score (mF1) of 0.77. Notably, DAEiS-Net has the lowest parameter count (15.3 M) among the compared models, indicating high efficiency. Qualitatively, visual comparisons showed DAEiS-Net outperforming other methods, particularly in accurately segmenting stains with complex boundaries and in challenging lighting conditions. The ablation study confirmed that each of the proposed modules (DAM, EISM, SPD) contributes significantly to the overall performance gains.

Limitations and Future Work

The authors acknowledge that while DAEiS-Net demonstrates strong performance, future work will focus on further optimizing the model, improving computational efficiency, and exploring its application to other tunnel diseases and segmentation tasks. Expanding the TWS dataset with more diverse tunnel environments and conditions is also suggested to enhance the model's robustness. The feature visualization experiments also suggest that the model's attention distribution varies with stain size, indicating a potential area for further investigation into how the model handles very small or very large stains.

Summary updated

Abstract

Tunnel disease detection and maintenance are critical tasks in urban engineering, and are essential for the safety and stability of urban transportation systems. Water stain detection presents unique challenges due to its variable morphology and scale, which leads to insufficient multiscale contextual information extraction and boundary information loss in complex environments. To address these challenges, this paper proposes a method called Deep Aggregation Network with Edge Information Supplement (DAEiS-Net) for detecting tunnel water stains. The proposed method employs a classic encoder–decoder architecture. Specifically, in the encoder part, a Deep Aggregation Module (DAM) is introduced to enhance feature representation capabilities. Additionally, a Multiscale Cross-Attention Module (MCAM) is proposed to suppress noise in the shallow features and enhance the texture information of the high-level features. Moreover, an Edge Information Supplement Module (EISM) is designed to mitigate semantic gaps across different stages of feature extraction, improving the extraction of water stain edge information. Furthermore, a Sub-Pixel Module (SPM) is proposed to fuse features at various scales, enhancing edge feature representation. Finally, we introduce the Tunnel Water Stain Dataset (TWS), specifically designed for tunnel water stain segmentation. Experimental results on the TWS dataset demonstrate that DAEiS-Net achieves state-of-the-art performance in tunnel water stain segmentation.
BibTeX
@inproceedings{wang2024daeis,
  author = {Yuliang Wang and Kai Huang and Kai Zheng and Shuliang Liu},
  title = {{DAEiS-Net: Deep Aggregation Network with Edge Information Supplement for Tunnel Water Stain Segmentation}},
  booktitle = {Sensors},
  year = {2024},
  url = {https://openreview.net/forum?id=hr0inSGKgm},
}
Figure