Multi-Classifier for RC Bridge Defects: Automated Concrete Inspection

A multi-source dataset of concrete defect images supports the development of automated bridge inspection systems. The dataset combines self-collected image samples with several Departments of Transportation inspection databases. It is used to train a three-stage deep neural classifier that can multi-classify defect types with an average mean score of 85%.

Use Cases

• Automated defect classification based on visual inspection images.
• Training objective inspection models to reduce human subjectivity in assessments.
• Benchmarking computer vision algorithms for conformity with bridge inspection guidelines.
• Supporting predictive maintenance models for ageing bridge infrastructure.

Strengths

• Dataset is multi-source, combining self-collected images with official Department of Transportation databases.
• The associated classifier model achieves an average mean score of 85% on defect classification.
• Dataset is associated with research from the University of Cambridge, indicating academic rigor.

Limitations

• Key metadata such as row count, column names, specific image count, and last update date are unknown across all sources.
• The dataset's specific composition and balance of defect types are not described.
• Relies on inference for data type; explicit confirmation of image format and annotations is missing.

Provenance

Source

University of Cambridge (Philipp Hüthwohl), combined with self-collected samples and several U.S. Departments of Transportation inspection databases.

Collection Method

Images were collected from bridge inspections, likely following existing inspection guidelines.

Geography

Likely includes data from the United States (via DoT databases) and other unspecified locations from self-collection. The research references the 2018 Genoa, Italy bridge collapse as context.