DAU-Net: A Dual-Attentive U-Net for Enhanced Semantic Segmentation in Underground Infrastructure Inspection

We introduce dual attentive U-Net (DAU-Net), a deep learning model for semantic segmentation designed to tackle the challenges of multiclass imbalanced datasets with limited samples. Our model processes video frames captured by closed-circuit television (CCTV) cameras to segment deficiencies in each frame. It captures both fine and broad details by integrating multiscale spatial–channel attention mechanisms. Specifically, the model utilizes a multiscale depth-separable block and an enhanced concurrent spatial–channel squeeze-and-excitation (EscSE) block to capture feature map dependencies more effectively. Our model design integrates various techniques, providing significant advantages over conventional methods, including reduced training time, more efficient feature extraction, and faster convergence. We evaluated DAU-Net on our real-world culvert-sewer defect dataset, which consists of infrastructure deficiencies captured under varying angles and lighting conditions. We also evaluated the model on two other datasets: the structural defect dataset and the benchmark cell nuclei dataset. Our model achieves state-of-the-art mean intersection over union (IoU) scores of 76.23%, 63.98%, and 87.31% on these datasets, respectively, surpassing previous methods by 26.59%, 12.97%, and 23.27%, on average. DAU-Net provides a robust solution for semantic segmentation in multiclass problems with limited training data. It excels in the critical domain of underground infrastructure inspection, where traditional manual methods are not only hazardous, time-consuming, and error-prone, but also pose significant risks to public safety. Our evaluation shows that DAU-Net is effective for both underground infrastructure inspection and biomedical image segmentation, proving its broad applicability to various semantic segmentation tasks. The code for our models is publicly available at https://tinyurl.com/DAUNet