AI-Based Culvert-Sewer Inspection

TL;DR

The paper addresses automated defect segmentation in culverts and sewer pipes under data scarcity by combining data-centric and model-centric solutions. It introduces a preprocessing pipeline with traditional augmentation and Dynamic Label Injection, a parameter-efficient FORTRESS architecture that integrates depthwise separable convolutions with adaptive Kolmogorov–Arnold Networks and multi-scale attention, and a few-shot semantic segmentation framework based on bidirectional prototypes and attention. Empirical results on the Culvert Sewer Defect Dataset show that augmentation plus DLI yields notable improvements across IoU and F1 metrics, while FORTRESS achieves state-of-the-art performance with far fewer parameters and lower FLOPs, and the few-shot approach demonstrates strong generalization for unseen defect classes. The work demonstrates practical impact by enabling accurate, real-time, data-sparse defect inspection, offering robust options for model deployment in real-world infrastructure monitoring. All methods are validated on a realistic dataset with class-imbalance and multi-scale challenges, highlighting their potential for scalable, autonomous inspections in critical water-management systems.

Abstract

Culverts and sewer pipes are critical components of drainage systems, and their failure can lead to serious risks to public safety and the environment. In this thesis, we explore methods to improve automated defect segmentation in culverts and sewer pipes. Collecting and annotating data in this field is cumbersome and requires domain knowledge. Having a large dataset for structural defect detection is therefore not feasible. Our proposed methods are tested under conditions with limited annotated data to demonstrate applicability to real-world scenarios. Overall, this thesis proposes three methods to significantly enhance defect segmentation and handle data scarcity. This can be addressed either by enhancing the training data or by adjusting a models architecture. First, we evaluate preprocessing strategies, including traditional data augmentation and dynamic label injection. These techniques significantly improve segmentation performance, increasing both Intersection over Union (IoU) and F1 score. Second, we introduce FORTRESS, a novel architecture that combines depthwise separable convolutions, adaptive Kolmogorov-Arnold Networks (KAN), and multi-scale attention mechanisms. FORTRESS achieves state-of-the-art performance on the culvert sewer pipe defect dataset, while significantly reducing the number of trainable parameters, as well as its computational cost. Finally, we investigate few-shot semantic segmentation and its applicability to defect detection. Few-shot learning aims to train models with only limited data available. By employing a bidirectional prototypical network with attention mechanisms, the model achieves richer feature representations and achieves satisfactory results across evaluation metrics.

Figures (17)

• Figure 1: Visualization of different computer vision tasks.
• Figure 2: Although direct connections in a CNN are sparse, there are multiple indirectly connected elements that increase the receptive field for each subsequent layer.
• Figure 3: Visualization of the U-Net architecture where the input image is downsampled to 32x32 pixels in the contracting path. The blue boxes refer to multi-channel feature maps whereas the white ones represent the copied feature maps from the contracting path to the expanding one. The number of channels are denoted on the top of the feature map and its size is stated inside or on its left ronneberger2015u.
• Figure 4: Visualization of the FPN architecture. The blue rectangles represent the feature maps at different pyramid levels. At each level, a small Multi-Layer Perceptron (MLP) is applied to sliding windows of size $5 \times 5$, shown as green boxes, to generate dense object proposals, each producing a segmentation mask of size $14 \times 14$, visualized in grey. The orange rectangles indicate the size of the image region that each proposal corresponds to. The light orange areas represent the image region size, while the dark orange areas visualize the typical object size at each pyramid level. This shows how deeper pyramid levels consider larger receptive fields lin2017feature.
• Figure 5: Visualization of the ViT architecture. The input image is divided into multiple image patches. For each patch a linear embedding with an added position embedding is handed over to the transformer encoder to create a rich feature representation. The transfomer encoder consists of multiple alternating multi-head self-attention and MLP blocks. An MLP head creates the final class predictions dosovitskiy21.