Speaker
Description
DGN: A DWT-Guided Frequency-Spatial Dual-Domain Dehazing Network for Sewer Inspection Images
Keywords: Discrete Wavelet Transform; dehazing; Sewer systems;
Abstract
Introduction. While spatial domain dehazing methods demonstrate effectiveness in removing concentrated haze, their limited capacity for holistic image restoration prompts exploration of frequency domain approaches. This paper proposes a Discrete Wavelet Transform-guided Network (DGN) to address structural-aware dehazing in sewer images. The method establishes enhanced structure awareness through two key mechanisms: (1) A wavelet attention module that autonomously weighs decomposed frequency components, particularly amplifying high-frequency features associated with water-pipewall borderlines; (2) A contrastive regularization framework in the frequency domain to better enhance the convergence speed of the network. Experimental results demonstrate superior performance with 146 in mean square error (MSE), 27.43 in peak signal-to-noise ratio (PSNR), and 0.9154 in structural similarity index measure (SSIM), achieved with 35.83M parameters. The proposed frequency-domain method provides an effective alternative solution for comprehensive sewer image restoration when dealing with complex haze distributions.
Material and methods. Building upon prior work in spatial domain dehazing that utilizes multi-task learning to enhance structural awareness of water boundary lines (Xia et al. 2022), this study extends the framework to the frequency domain through Discrete Wavelet Transform (DWT). The proposed DWT-Guided Network (DGN) introduces a dual-domain architecture that concurrently processes spatial and frequency features to improve structural preservation in pipeline defect detection. Haar wavelet decomposition segregates input features into low-frequency (LL) and high-frequency components (LH, HL, HH), where the latter include critical edge features such as water boundaries. These high-frequency subbands are selectively amplified through novel DWT-Guided Residual Blocks (DGRBs), which implement a wavelet attention mechanism to prioritize structural details while suppressing redundant spatial information.
Each DGRB operates by decomposing input features into four frequency subbands via DWT, followed by concatenation and upsampling of the high-frequency components$\ f_{H}$ to restore spatial resolution.
$f_{H} = Upsample\left( \text{Concate}\left( f_{\text{LH}},\ f_{\text{HL}},\ f_{\text{HH}} \right) \right)$ (1)
A sequence of convolutional layers and a sigmoid activation then refine these signals before re-integrating them with spatial features processed through coordinate attention modules (Hou, Zhou, & Feng 2021). This fusion strategy ensures simultaneous retention of spatial context and enhancement of edge-specific frequency features.
$f_{O} = f + f_{f} + f_{s}$ (2)
To enforce frequency-domain consistency during optimization, DWT-Guided Contrastive Regularization (DGCR) is introduced, establishing a triplet relationship among dehazed outputs (anchors), hazy inputs (negatives), and ground-truth images (positives). By projecting these triplets into the DWT-transformed frequency space, DGCR minimizes the spectral divergence between anchors and positives while maximizing separation from negatives. This regularization constrains the solution space in both spatial and frequency domains, effectively addressing the inherent ambiguity in dehazing tasks.
The network architecture comprises sequential convolutional layers and 11 DGRBs, with each block performing simultaneous channel-wise feature extraction and wavelet-based frequency modulation. Implementation employs Haar wavelets for their computational efficiency in generating four subbands, coupled with convolutions to maintain structural coherence. The combination of wavelet attention and contrastive regularization demonstrates superior capability in recovering edge details under varying haze densities compared to conventional spatial-domain approaches.
Results and discussion. Using the S2B drainage pipe defect dataset, DGN demonstrates superior image recovery with notable parameter efficiency. As shown in Table 1, our method achieves the best SSIM (0.9154) and PSNR (27.43 dB) while maintaining 35.83M parameters. The 0.15 dB PSNR improvement over SANL-Net confirms that frequency-aware processing better preserves structural details critical for defect inspection.
The quantitative superiority stems from two aspects: 1) Separate processing of low/high-frequency components reduces entropy interference, lowering MSE to 146 (-0.68% vs. SANL-Net); 2) Wavelet domain constraints enhance feature discrimination, improving SSIM by 2.08% compared to DGNL-Net. Though marginally larger than specialized dehazing networks (35.83M vs. 15.47M in SANL-Net), our parameter scale remains practical for industrial deployment -- 74.5% smaller than MSBDN-DFF while delivering better performance.
Table 2, 3 and 4 show the performance improvement of the images in downstream tasks of pipeline damage detection after DGN dehazing. The segmentation and localization tasks are conducted on the Pipe (Pan, Zheng, & Guo 2020) dataset, while the classification task is performed on the Sewer-ML (Haurum et al. 2021) dataset. It can be observed that in downstream defect analysis tasks, the wavelet attention architecture significantly enhances key detection metrics: a 10.41% increase in mIoU for JO class in semantic segmentation, a 17.59% improvement in mAP for IN localization with YOLOv5s, and a 9.99% growth in F1-score for FS detection in classification tasks.
Conclusions. This study proposes a DWT-Guided Network (DGN) framework based on Discrete Wavelet Transform (DWT), which effectively addresses the challenges of texture suppression and detail distortion in pipeline defect detection through dual-domain guidance in frequency and spatial domains. The network utilizes a DWT module to decompose input features into low-frequency and high-frequency components. In the spatial domain, the coordinate attention method preserves spatial information, and the input features are fused with frequency-domain and spatial-domain features for final output. The Domain-Guided Contrastive Regularizer (DGCR) employs a contrastive learning approach to pull closer positive feature pairs in the frequency domain while pushing apart negative pairs. Experimental results demonstrate that DGN achieves 0.9154 SSIM in pipeline image restoration tasks under haze interference, representing a 2.13% improvement over traditional spatial-domain methods. In pipeline damage detection tasks such as segmentation, localization, and classification, our DGN network significantly improved the scores of hazing images: a 10.41% increase in mIoU for JO class in semantic segmentation, a 17.59% improvement in mAP for IN localization with YOLOv5s, and a 9.99% growth in F1-score for FS detection in classification tasks. These advancements validate the importance of DWT domain modeling for pipeline defect detection, establishing a novel technical paradigm for intelligent inspection of drainage systems.
