A Unified Computer Vision Framework for Automated Crack Detection and Quantification in Reinforced Concrete Beams

This repository supports the research titled “A Unified Computer Vision Framework for Automated Crack Detection and Quantification in Reinforced Concrete Beams.” The study proposes a fully automated, non-contact framework for structural crack detection, segmentation, and quantitative measurement using computer vision and deep learning, integrated with calibrated stereo vision.The primary objective of the research is to bridge the gap between visual crack detection and engineering-scale crack quantification required for Structural Health Monitoring (SHM). Unlike conventional image-based approaches that remain limited to 2D detection or qualitative assessment, the proposed framework delivers metric crack parameters—including crack width, length, orientation, and initiation location—directly in physical units (millimetres).The framework integrates multiple deep learning models within a single pipeline:Beam support condition classification using a lightweight CNN to identify structural boundary conditions.Beam semantic segmentation and corner-point extraction to isolate the structural element and enable geometric calibration.Crack detection using CNN-based classifiers to localize crack regions.Crack semantic segmentation using a U-Net architecture for pixel-level crack delineation.Stereo vision–based pixel-to-millimetre conversion, eliminating the need for physical scale markers.Automated crack parameter quantification, including:Crack length via skeletonization and regression fittingCrack width using Euclidean distance transformsCrack orientation using K-means clustering and linear regressionCrack initiation point localization relative to beam geometryThe dataset provided in this repository includes annotated images, segmentation masks, and calibration-related data used to train and validate the models. Experimental validation was conducted on reinforced concrete beam specimens under controlled laboratory loading conditions. The framework achieved sub-millimetric crack width accuracy, millimetric crack length accuracy, sub-degree orientation error, and high spatial accuracy in crack initiation point detection, demonstrating its suitability for engineering-grade SHM applications.This dataset and accompanying methodology are intended to support:Reproducibility of the reported resultsBenchmarking of crack detection and quantification algorithmsFurther research on vision-based structural health monitoringExtension to other reinforced concrete elements such as slabs, columns, and walls