@article{Zhu2026, 
author = {Jiasong Zhu and Antai Chen and Wenyu Jiang and Yuansheng Hua and Siyu Chen and Qingquan Li},
title = {EDR-Fra: an enhanced defect recognition framework for sewer floating capsule robots},
year = {2026},
journal = {Geo-Spatial Information Science},
volume = {29},
number = {4},
pages = {3192-3213},
keywords = {Sewer pipeline defect recognition, capsule robots, defogging and deblurring, deep learning, infrastructure health monitoring, object detection, image quality enhancement},
url = {https://www.sciopen.com/article/10.1080/10095020.2025.2610866},
doi = {10.1080/10095020.2025.2610866},
abstract = {The safe operation and maintenance of urban water supply and drainage systems are crucial for sustainable socio-economic development. Traditional pipeline health inspection methods suffer from complicated operation, high cost, and low efficiency. To tackle these issues, the sewer floating capsule robot (SFCR) is proposed as a new and convenient automated inspection solution. However, sewer pipelines are often narrow, and water flow is rapid, which significantly degrades the quality of SFCR data, leading to issues such as uneven lighting, water vapor fogging, and motion blurring. To address these problems, a computer-based enhanced defect recognition framework (EDR-Fra) is proposed to improve image qualities and defect recognition performance of SFCRs. The framework consists of three core networks: FogClearNet (FoCNet), CheckerboardClearGAN (CCGAN), and YOLOv5. FoCNet employs a dual-branch feature fusion module, integrating channel-level and global-level high-order features to effectively eliminate water vapor fogging influence on images. CCGAN is built on the architecture of generative adversarial networks (GANs), where the generator produces deblurred images that are then fed to the discriminator for quality evaluation. YOLOv5 is employed to identify and localize various types of sewer defects (i.e. crack, break, and deformation). Experimental results on the SFCRs dataset show that FoCNet improves the image peak signal-to-noise ratio (PSNR) by 24.04 and structural similarity (SSIM) by 0.271. CCGAN improves PSNR by 6.33 and SSIM by 0.066. After image enhancement, the defect recognition performance of YOLOv5 increases nearly 12% in mean average precision (mAP) at the speed of 113.12 fps. These results demonstrate that the proposed framework EDR-Fra is a promising tool for efficient and low-cost defect detection in urban underground sewer pipelines.}
}