Metricéa Lab 380: European field image dataset for concrete crack classification
Description
This dataset, collected by Metricéa Lab, comprises 380 field images of concrete civil infrastructure, specifically curated to address the "domain shift" phenomenon in deep learning-based Structural Health Monitoring (SHM). While many existing concrete crack datasets are generated in controlled laboratory environments or scraped from the internet, these images were captured on-site during operational bridge and infrastructure inspections in the Île-de-France region. The image acquisition aligns with the rigorous criteria of the French IQOA (Image de la Qualité des Ouvrages d'Art) structural evaluation framework. The dataset is designed for binary classification tasks (Cracked vs. Uncracked) and captures the complex photometric and topological artifacts routinely encountered in real-world civil engineering. This includes variable natural lighting, shadows, surface weathering, and complex concrete textures that cause standard laboratory-trained AI models to fail in the field. This dataset has been empirically proven effective for domain adaptation: it was successfully utilized to fine-tune a pre-trained ResNet50 architecture, demonstrating that a highly constrained, field-specific dataset of only 380 images is sufficient to restore diagnostic accuracy from a degraded 14.3% (when using a laboratory baseline) to 100% on operational field data.
Files
Steps to reproduce
Steps to reproduce (Data Acquisition and Processing): Phase 1: Field Data Acquisition on-site inspection: Raw images were physically captured during operational structural inspections of concrete civil infrastructure located in the Île-de-France region. Regulatory Framework: The photographic acquisition was conducted following the rigorous guidelines of the French IQOA (Image de la Qualité des Ouvrages d'Art) standard, ensuring the capture of authentic field conditions (complex lighting, shadows, and natural surface weathering). Manual Curation: The collected photographs were expertly reviewed, cropped to isolate the structural elements, and strictly categorized into two definitive classes: Cracked and Uncracked. Phase 2: Computational Usage for Domain Adaptation4. Data Loading: Extract the dataset into a standard ImageFolder directory structure.5. Preprocessing: Resize the images to 224 x 224 pixels and apply ImageNet normalization to match standard ResNet50 input requirements.6. Fine-Tuning: Use these field images to perform selective layer unfreezing (e.g., layers 3 and 4 of a pre-trained ResNet50 model) to bridge the domain shift gap between laboratory environments and operational reality.