Prototypical Learning with Attention Enhancement and E-FPN for Few-Shot Segmentation of Culvert and Sewer Defects

Underground infrastructure inspection-especially for culverts and sewers-faces unique challenges: manual inspection is laborious, time-consuming, and risky for workers, while automated detection relying on deep learning often struggles with scarce labeled data and the need to adapt to new defect types. Few-shot semantic segmentation emerges as a key solution to these issues, yet existing frameworks still fall short: they demand large-scale annotated datasets to perform well, and fail to effectively learn and segment novel defect categories when only a handful of samples are available. To fill this gap, this paper proposes a dedicated few-shot semantic segmentation framework, which fuses the Enhanced Feature Pyramid Network (E-FPN) with prototypical learning and attention mechanisms, specifically optimized for culvert and sewer defect detection. The framework's three key innovations are as follows: (1) An E-FPNbased adaptive encoder. By integrating InceptionSepConv blocks and depth-wise separable convolutions, this encoder efficiently extracts multi-scale features-capturing both micro-scale cracks and small surface defects, as well as large-scale structural failures-addressing the variable size of infrastructure defects in real-world scenarios. (2) A prototypical learning module equipped with masked average pooling. This module generates reliable class prototypes from limited support samples by filtering out background noise to focus on defect-centric features, ensuring effective representation of each defect category even with minimal training data. (3) A multi-modal attention mechanism. Combining global self-attention, local self-attention, and crossattention, this mechanism refines feature embeddings, enhances the distinguishability between defect types and backgrounds, and improves the accuracy of prototype matching for query image segmentation. Comprehensive tests were conducted on a challenging culvert-sewer defect dataset (characterized by severe class imbalance, with common defects overrepresented and critical rare defects underrepresented). Results show the framework delivers strong few-shot performance: when trained under an 8-way 5-shot configuration, it achieves an F1-score of 82.55This framework directly addresses the core need of infrastructure inspection systems: rapid adaptation to new defect types with limited additional labeled data. It not only reduces the cost and effort of data annotation for underground infrastructure detection but also enables more flexible, efficient maintenance planning for critical culvert and sewer systems-laying a foundation for wider adoption of AI-driven automated inspection in complex underground environments. Supplementary Material File (manuscript5 (2).pdf) - Download - 291.94 KB Information & Authors Information Version history Copyright This work is licensed under a Non Exclusive No Reuse License.

Authors Metrics & Citations Metrics Article Usage 222views 114downloads Citations Download citation Francesco Fabiano, Takahiro Otani. Prototypical Learning with Attention Enhancement and E-FPN for Few-Shot Segmentation of Culvert and Sewer Defects. Authorea. 17 November 2025. DOI: https://doi.org/10.22541/au.176341002.24765995/v1 DOI: https://doi.org/10.22541/au.176341002.24765995/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu.