Data-Driven Detection and Evaluation of Damages in Concrete Structures: Using Deep Learning and Computer Vision
Saeid Ataei, Saeed Adibnazari, Seyyed Taghi Ataei · 2025
Read Data-Driven Detection and Evaluation of Damages in Concrete Structures: Using Deep Learning and Computer Vision by Saeid Ataei, Saeed Adibnazari, Seyyed Taghi Ataei (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-11836.txt.
Introduction
Saeid Ataei, Saeed Adibnazari, and Seyyed Taghi Ataei trained AI to spot damage in concrete in a 2025 study. Cracks and spalls in bridges, tunnels, and walls are now found by eye, a slow path prone to error. The team asked whether deep learning on photos could do the job faster and with fewer misses.
They grew a set of four hundred photos into near eleven thousand with flips and color shifts. Two top models learned to outline each flaw. The YOLO model scored higher and ran at forty frames per second, twice the speed of its rival. The authors back it for live checks on site, and the slower model for deep offline reads.
Catalog record
| CREATOR | Saeid Ataei, Saeed Adibnazari, Seyyed Taghi Ataei |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-11836
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Data-Driven Detection and Evaluation of Damages in Concrete Structures: Using Deep Learning and Computer Vision Saeid Ataei, Saeed Adibnazari, Seyyed Taghi Ataei Abstract Structural integrity is vital for maintaining the safety and longevity of concrete infrastructures such as bridges, tunnels, and walls. Traditional methods for detecting damages like cracks and spalls are labor-intensive, time-consuming, and prone to human error. To address these challenges, this study explores advanced data-driven techniques using deep learning for automated damage detection and analysis. Two state-of-the-art instance segmentation models, YOLO-v7 instance segmentation and Mask R-CNN, were evaluated using a dataset comprising 400 images, augmented to 10,995 images through geometric and color-based transformations to enhance robustness. The models were trained and validated using a dataset split into 90% training set, validation and test set 10%. Performance metrics such as precision, recall, mean average precision ([email protected]), and frames per second (FPS) were used for evaluation. YOLO-v7 achieved a superior [email protected] of 96.1% and processed 40 FPS, outperforming Mask R-CNN, which achieved a [email protected] of 92.1% with a slower processing speed of 18 FPS. The findings recommend YOLO-v7 instance segmentation model for real-time, high-speed structural health monitoring, while Mask R-CNN is better suited for detailed offline assessments. This study demonstrates the potential of deep learning to revolutionize infrastructure maintenance, offering a scalable and efficient solution for automated damage detection. K eywords Concrete Structures \cdot Data-Driven \cdot Deep Learning \cdot Damage Detection \cdot YOLO-v7 \cdot Mask R-CNN \cdot Instance Segmentation 1 Introduction Structural health monitoring (SHM) is essential for ensuring the safety, efficiency, and longevity of infrastructure. Periodic inspections are particularly critical for concrete structures, such as bridges, building walls, tunnels, and dams, to identify potential defects that may compromise structural integrity. Common indicators of structural health deterioration include defects like cracks and spalls [ 4 ] . These issues often arise due to fatigue and periodic loading, underscoring the need for early, timely, and automated diagnostic methods [ 5 ] . Implementing proactive measures can significantly reduce financial costs and mitigate human safety risks associated with structural failures. Concrete is among the most widely used construction materials due to its affordability, versatility, and extensive applications in civil works and infrastructure development. Consequently, advancements in monitoring and maintaining its structural integrity are paramount. The rise of data-driven methods for analyzing and predicting structural defects represents a paradigm shift from traditional manual inspections [ 8 ] . These methodologies, grounded in verifiable data analysis rather than intuition or subjective judgment, provide a robust framework for deriving insights and developing solutions [ 9 ] . Such methods enable predictive assessments by leveraging knowledge from historical and current data, thereby reducing biases and false assumptions. Recent advancements in machine learning and artificial intelligence, particularly deep learning, have transformed SHM processes. Vision-based deep learning algorithms can efficiently identify structural defects, including cracks and spalls, with superior accuracy compared to manual inspections. These automated methods expedite the diagnosis process, decrease reliance on specialized expertise, and prevent potential human errors [ 11 , 12 , 13 ] . By automating the detection process, deep learning reduces time and costs while enhancing the reliability of defect identification [ 14 ] . 1.1 Problem Definition Data-driven techniques for defect detection in concrete structures are gaining traction due to advancements in hardware and software technologies. These methods can be categorized into two groups: 1. Non-intelligent Monitoring: Traditional approaches, such as visual inspections and manual experiments, heavily depend on human expertise, leading to safety risks, inefficiencies, and inaccuracies. These methods are often labor-intensive and time-consuming [ 15 ] . 2. Intelligent Monitoring: Advanced methodologies, including image processing, machine learning, and deep learning, address the limitations of traditional approaches. These techniques facilitate cost-effective, precise, and automated monitoring of structural health [ 16 ] . Concrete SHM often involves managing large datasets. Techniques like instance segmentation in deep learning help in extracting meaningful patterns from this data. By employing vision-based algorithms, structural damage can be accurately identified, classified, and localized, ensuring timely interventions. Moreover, the ability to utilize pre-trained models and transfer learning has significantly enhanced the accessibility and efficiency of deep learning-based SHM systems [ 17 ] . 1.2 Methods for Detection Deep learning, a subset of machine learning, encompasses various learning paradigms, including supervised, semi-supervised, and unsupervised learning [ 18 ] . These approaches require collecting, preprocessing, and preparing data before applying algorithms for classification and pattern extraction. Techniques such as transfer learning, augmented datasets, and pre-trained networks (e.g., the SDNE2018 dataset, which contains 56,000 images of bridges and walls) expedite the training process and improve accuracy [ 19 ] . Classification and segmentation methods are crucial in SHM. Classification involves categorizing data into distinct classes, such as cracks, spalls, or intact regions. Instance segmentation extends classification by identifying individual defects within a broader context, providing detailed insights into structural health [ 20 ] . Additionally, few-shot learning methods address data limitations by focusing on learning patterns from minimal data samples. This innovative approach has proven effective across diverse fields, including image classification, natural language processing, and structural damage detection [ 21 ] . 1.3 State of the Research Field Concrete defect detection has evolved from heuristic and machine vision techniques to sophisticated deep learning methods. Traditional image processing filters, such as Canny and Sobel, were once prevalent but limited in handling complex structural defects [ 10 ] . The advent of artificial neural networks (ANNs) brought significant advancements, with convolutional neural networks (CNNs) leading the charge [ 11 ] . CNN-based techniques bypass manual feature extraction, enabling direct learning of image features. For instance, Cha and Choi [ 20 ] demonstrated the efficacy of deep CNN models under variable brightness and shadow conditions. Similarly, studies employing hybrid models—merging CNNs with traditional methods like Otsu filtering—achieved higher accuracy in detecting cracks and predicting their direction [ 21 ] . Recent research also explores transformer-based models, which, despite higher computational costs, exhibit enhanced flexibility and reduced bias compared to CNNs [ 28 ] . Several pioneering works have focused on classification and segmentation. Noushin Bigdeli et al. [ 15 ] utilized a CNN-based approach to classify concrete cracks into distinct categories, achieving an impressive 99.3% accuracy. Other studies employed advanced pre-trained networks, such as ResNet and AlexNet, for transfer learning, significantly improving defect detection precision [ 17 , 26 ] . For instance, the ResNet-50 model—renowned for its depth and accuracy—has been instrumental in concrete damage classification [ 25 ] . Object detection algorithms like YOLO and R-CNN are gaining momentum for their ability to localize and categorize structural defects. Zhou et al. [ 26 ] proposed a faster R-CNN model to identify and locate seismic damages in reinforced concrete columns with an 80% accuracy. Similarly, Maedeh et al. [ 27 ] used single-stage object detection networks, such as SSD-MobileNet, to identify road damages with a 75% accuracy. 1.4 Diverging Hypotheses and Controversies While deep learning techniques promise unprecedented accuracy, there are debates surrounding their scalability and computational efficiency. Transformer-based models, despite their high potential, pose challenges due to elevated computational demands. Moreover, data-driven approaches often rely on the availability of large, high-quality datasets. Inconsistent data labeling and variability in imaging conditions further complicate model generalization [ 28 ] . Another