Issuu

AI-Driven Crack Detection for Civil Infrastructure Safety

Amarsinh B. Landage1 , Pranjali D. Patil2, Rajas P. Kulkarni3 , Vedant S. Gokhale4

Assistant Professor, Department of Civil and Infrastructure Engineering, Government College of Engineering, Ratnagiri,415612, India,

Research Scholar, Department of Civil and Infrastructure Engineering, Government College of Engineering, Ratnagiri, 415612, India

Abstract-The rapid deterioration of civil infrastructure due to ageing, overloading, and environmental stress demands automated, reliable, and intelligent structural health monitoring solutions. Conventional crack detection practices relying on manual visual inspections are timeconsuming, subjective, and unsafe for inaccessible structures.

This study presents Crack Detection using AI, an AI-driven crack detection framework integrating Convolutional NeuralNetworks(CNNs),transferlearningviaMobileNetV2, Computer Vision, and Explainable AI (Grad-CAM) for automated classification and severity assessment of structural cracks from digital images. A multi-source dataset of 3,420 images collected from mobile devices, UAV/drone platforms, and public datasets (SDNET2018, Crack500) was assembled, preprocessed, and augmented to 10,260training-readyimages.

The MobileNetV2-based architecture fine-tuned via twophasetransferlearningachieved96.84%testaccuracy,with precision of 95.72%, recall of 97.10%, F1-score of 96.40%, and AUC-ROC of 0.983. The system is deployed as a publicly accessible web application with Grad-CAM heatmap visualizations and automated PDF report generation. Results demonstrate a 25–30% accuracy improvement over manual inspection with over 100× speed gain, confirming the viability of AI for scalable, standards-compliant structuralinspectioninIndiancivilengineeringpractice.

Keywords-Crack Detection, Convolutional Neural Networks, Transfer Learning, Grad-CAM, Structural Health Monitoring, Computer Vision.

1. INTRODUCTION

The exponential growth of urbanization, ageing infrastructure, and increased service loads on civil structures has significantly intensified the global demand for reliable, automated, and intelligent structural health monitoring (SHM) systems. Structural cracks are among the earliest signs of distress in concrete infrastructure, resulting from overloading, thermal stresses, foundation settlement,shrinkage,ormaterialfatigue(Sinha&Fieguth, 2006; Mohan & Poobal, 2018). Early detection and timely

repair of cracks are essential to prevent progressive deterioration,ensureserviceability,andavertcatastrophic structuralfailures.

Conventionalcrackdetectionreliesheavilyonmanual visual inspection, where trained engineers physically examine structural surfaces for defects. While simple and inexpensive, this approach is time-consuming, labourintensive,subjective,andinherentlyunsafefortall,remote, or underground structures such as bridges, elevated highways, and tunnel linings. The reliability of inspection outcomes depends directly on inspector experience and attentiveness, leading to inconsistent and error-prone assessments. With the rapid advancement of Artificial Intelligence (AI) and Computer Vision (CV), automated crack detection using deep learning has emerged as a transformativesolution.

AI-driven crack detection integrates Deep Learning (DL), Convolutional Neural Networks (CNNs), Grad-CAM explainability, and Unmanned Aerial Vehicles (UAVs) to enhance the speed, accuracy, and reliability of crack identification(Zhangetal.,2016;Chaetal.,2017).Existing research demonstrates promising results using CNN architectures for binary crack classification (Shetty et al., 2019),pavementcrackdetection(Mane&Patil,2019),and real-time object detection using YOLOv4 (Rajesh et al., 2024). However, most studies examine these technologies individually rather than within a fully integrated, deployable inspection framework tailored to Indian field conditions.

Although extensive research exists on CNN-based crackclassification,Grad-CAMexplainability,UAV-assisted imaging, and real-time deployment, most studies examine thesecomponentsinisolation.Thereisaclearlackofwork combining multi-source data acquisition, robust preprocessing, deep learning classification, explainability, and production deployment into a unified inspection system. This gap identifies the critical need for a comprehensive AI-driven crack detection framework whichthepresentstudy,CrackDetectionUsingAI,aimsto address.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

Table 1: Comparison between Conventional and AIBased Crack Detection Systems

Parameter ConventionalIn spection

Crack Detection using AI(AIBased)

Accuracy 65–75% (subjective) 96.84% (deterministic)

Speed Hours per structure <2secperimage

Consistency Variable (inspectordependent)

Accessibility Accessible surfacesonly

Explainability Expertjudgment (implicit)

Consistent, fatigue-free

Mobile + UAV + Web

Grad-CAM visual proof

Reporting Manual notes / photos Automated PDF reports

2. METHODOLOGY

The methodology of Crack Detection usingAI outlines thesystematicframeworkadoptedtodesign,develop,and evaluate an AI-driven crack detection system. A hybrid research approach combining both qualitative and quantitative methodologies is adopted, where the qualitative component includes literature synthesis, dataset curation protocols, and identification of critical preprocessing parameters, while the quantitative component focuses on data-driven CNN model development, augmentation pipelines, AI prediction accuracy,systemsimulation,anddeploymentperformance evaluation. This establishes a structured pathway integrating data acquisition, deep learning, explainability, and web deployment into a unified workflow. The framework begins with raw image input from multiplatform sources and terminates with actionable structural health outputs validated through experimental evaluationundervariousreal-worldinspectionscenarios.

Image 1: Methodology Flowchart

Crack Detection using AI — Methodology Flowchart (START)

▼SENSORDATACOLLECTION·Collectsrawimages (mobilecamera,UAV/drone,publicdatasets: SDNET2018,Crack500)

▼DATAPREPROCESSING·Resize224×224px, Normalize[0,1],GaussianFilter,CLAHEenhancement (+15%hairlinerecall)

▼AI/DEEPLEARNINGMODEL·MobileNetV2CNN TransferLearning,Fine-tuning(35epochstotal,2phase)

▼MODELTRAINING&VALIDATION·Adam optimizer,EarlyStopping,70/20/10train/val/test split;96.84%testaccuracy

▼GRAD-CAM&WEBDEPLOYMENT Heatmap visualization,FastAPIbackend,GitHubPages,PDF reportgeneration

▼(END)

Toacquirerepresentativetrainingdata,amulti-source datacollectionstrategywasadoptedusinghigh-resolution smartphone cameras (minimum 12 MP) for on-site image collection across reinforced concrete beams, columns, slabs, retaining walls, pavements, and building facades. Images were captured under varying distances (0.3–5 m), orientations, lighting conditions, and surface states (dry, wet,dusty,andweathered).UAVsweredeployedforaerial imaging of inaccessible structures including bridge decks and elevated RCC slabs at altitudes of 2–15 metres, providing geotagged nadir and oblique imagery. Public datasetsincludingSDNET2018,theConcreteCrackImages Dataset(40,000images),andCrack500wereincorporated to supplement field data and improve generalization acrossdiversestructuraltypes.

Data preprocessing transforms raw collected images into a clean, standardized, and model-ready dataset. All images were resized to 224×224 pixels and pixel intensities normalized to [0.0, 1.0]. Gaussian filtering removed high-frequency sensor noise while preserving crack boundary detail. CLAHE was applied to enhance crack visibility in shadow and low-light images, yielding a measured15%improvementinhairlinecrackrecall.Eight augmentation operations random rotation (±30°), horizontal and vertical flipping, brightness variation (0.7–1.3×), Gaussian noise injection, zoom/crop (10–20%), perspectivedistortion,andsheartransformationexpanded the3,420-imagedatasetto10,260training-readyimages,a 3×increase.

The core detection architecture is based on MobileNetV2, a lightweight CNN pre-trained on ImageNet, selectedforitsdepthwiseseparableconvolutions,inverted residual structure, and suitability for edge and mobile deployment. Transfer learning was conducted in two phases:Phase1frozeMobileNetV2layersandtrainedonly the custom classification head (Global Average Pooling → Dense256ReLU→ Dropout 0.5→ OutputSoftmax)for 15 epochs at lr=0.001. Phase 2 unfroze the top 30–50 MobileNetV2 layers for fine-tuning at lr=1e-5 for 20 additional epochs. Ensuring engineer confidence through continuous explainability involves Gradient-weighted Class Activation Mapping (Grad-CAM), which generates a

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

colour-coded heatmap overlaid on original images. Risk reduction and cost minimisation are executed by aggregating classification results, confidence scores, and Grad-CAM outputs into structured PDF inspection reports with severity grades (Minor / Moderate / Severe) per IS 456:2000.

Table 2: Methodological Framework and AI/CV Integration

ProjectOb jective

SensingTec hnology

CrackClass ification Mobile images,UAV imagery,SDN ET2018,Crac k500

Explainabi lity(XAI)

Preproces sing& Augmenta tion

CNNfinal convlayer gradients

AI/Machine LearningAlg orithm

PrimaryO utcome/S afety Action

MobileNetV2 CNN,Transfe r Learning,Sof tmaxoutput 96.84% accuracy;c onfidence score;5class output

GradCAM(Gradie ntweightedCla ss Activation) Visual heatmaps; 94.3% correctcra ck localizatio n

Raw 3,420 images(Mobi le + UAV +Public datasets) CLAHE, GaussianFilt er, TFImageDat aGen Dataset3× to10,260 images;+1 5% hairline recall

SeverityAs sessment Classified images;crack width &coverage area

RealTimeDeplo yment

Web image upload;FastA PI backend;TF.j smobile

Rule-based +IS 456:2000 /ACI318-19 Minor/Mo derate/Se vere grading;st andardscompliant

TensorFlow SavedModel; TF Lite; GitHub 1.8 sec API;42 FPS GPU;12 FPS mobile

3. RESULTS AND DISCUSSIONS

The implementation of the Crack Detection using AI framework yielded significant quantitative improvements across all five operational objectives, fundamentally transformingstructuralinspectionfroma reactivemanual process to a highly predictive automated discipline. For the initial objective of dataset development, AI-enhanced preprocessing drastically improved training data quality. Gaussian noise removal yielded a 12% improvement in

training stability, CLAHE enhancement improved annotation accuracy to 98.7%, and normalisation accelerated training convergence by approximately 20%. The augmentation pipeline expanded the original 3,420image dataset to 10,260 training-ready images. Building upon this, the MobileNetV2-based CNN trained via twophase transfer learning demonstrated 96.84% test accuracy, significantly outperforming traditional manual inspection (65–75%) and classical image processing approaches.

Themodelachievedstabletrainingaccuracyof97.2% and validation accuracy of 96.5% by epoch 28, with early stopping triggered due to no further improvement in validationloss.Traininglossconvergedfromaninitial0.68 to a final 0.09, indicating effective learning across all five crack categories. The Grad-CAM explainability layer demonstrated that 94.3% of true positive classifications showed correctly localized model attention on crack regions, confirming model transparency essential for engineeringacceptance.

Image 2: Predicted vs Actual Accuracy Curve

*Earlystoppingtriggeredatepoch28(val_lossplateau)

The most critical outcome of this research centres on the successful deployment of a centralised, AI-driven web platform designed to continuously classify structural images and dynamically visualize localized crack hazards with Grad-CAM overlays. Acting as the terminal graphical user interface, this web application processes uploaded images via a FastAPI backend, performs MobileNetV2 inference, generates Grad-CAM heatmaps, and delivers structuredPDFinspectionreportswithin1.8seconds.The core classification capability is driven by its five-class softmax output, which deterministically evaluates crack type and confidence score. Based on this real-time classification, the system automatically assigns severity gradesalignedwithIS456:2000thresholds.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net

Figure 3: Dashboard Output for Crack Classification and Severity Assessment

200

250

300

400

500

Analysisofthedashboardoutputrevealsprecisecrack classification performance. The initial 50–150 metre images maintained low risk scores confirming correct No Crack and Hairline classifications. As crack complexity increased including longitudinal bending cracks at 200 m, transverse overloading cracks at 250 m, and alligator fatigue networks from 300 m onwards the system escalated severity grades automatically. At the terminal 500-metre boundary, alligator crack coverage exceeding 35% with displacement above 1.2 mm produced a terminal HIGH RISK status (score 148.50), prompting generation of immediate inspection alerts and structural intervention recommendations per IS 456:2000 requirements.

Beyond the classification dashboard, Grad-CAM integration profoundly enhanced the practical utility of inspection output. By applying Grad-CAM to every classifiedimage,94.3%oftruepositivedetectionsshowed concentratedhigh-activationregionspreciselyalongcrack paths.Fifteenfalsenegativecasespredominantly involved hairline micro-cracks on heavily textured surfaces, while sixteen false positives were caused by deep surface stains and construction joint lines with sharp edge profiles similartocrack boundaries providingactionabledirection for targeted dataset augmentation. Finally, the web applicationdelivered1.8-secondAPIround-trip,42FPSon server GPU, and 12 FPS on mobile Android via TensorFlow.js, confirming practical deployment viability acrossalltargethardwareplatforms.

4. CONCLUSIONS

The present study successfully develops a comprehensive framework for AI-driven crack detection and structural safety by integrating CNN-based predictive analytics, Grad-CAM explainability, and end-to-end web application deployment. The research demonstrates that conventional inspection practices,whichprimarilyrely on manual visual judgment, can be significantly improved through intelligent automated technologies tailored to Indian civil engineering field conditions. The proposed model emphasises proactive hazard detection, enhanced inspection transparency, and optimised operational performance.

The results clearly indicate that AI-enhanced crack detection using MobileNetV2 transfer learning leads to substantial improvements in structural safety through high-fidelity classification and early severity warning mechanisms. The intelligent processing of multi-source training data minimises missed cracks and Grad-CAM provides visual verification of model decisions essential for professional engineering acceptance. Similarly, the web-based deployment proves an effective practical solution by predicting crack severity and ensuring IS 456:2000-compliant maintenance recommendations. The combined effect delivers improved inspection performance in terms of detection accuracy, engineer trust,andcostefficiency.

The study highlights the importance of adopting an integrated approach rather than deploying individual detection models in isolation. The combination of deep learning, explainability,and deploymentplatformscreates a synergistic effect that enhances overall structural health management. The proposed framework provides a practical and scalable solution applicable to buildings, bridges, pavements, and tunnels in geologically and environmentallydiverseIndianinfrastructure.

Table 3: Outcomes of the Study

Parameter Improvement Achieved DetectionAccuracy 96.84%

InspectionSpeed < 2 sec per image (100×+ vs.manual)

Grad-CAMCorrectFocus 94.3 % of true positive cases

HairlineCrackRecall + 15 % via CLAHE preprocessing

FalseNegativeRate 2.7 % (safety-critical miss rate)

StructuralCostSavings Early detection reduces repaircostby15%+

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

The findings contribute to the growing field of structural engineering and civil infrastructure inspection by providing a detailed and practical framework for AIbased safety management. The model aligns with global sustainableinfrastructuregoals(SDGs),particularlySDG9 and SDG 11, demonstrating the potential of integrating advanced digital technologies in civil engineering applications. It can serve as a reference for structural engineers, project managers, and policymakers in developing safer, resilient, and highly efficient infrastructureinspectionsystems.

5. FUTURE SCOPE

While the present study establishes a robust and comprehensive framework for AI-driven crack detection and structural safety, several avenues for future research and field implementation remain. The successful deployment of this predictive analytics system provides a strong foundation for fully autonomous, drone-integrated structural inspection missions. Future iterations can transitionfromprovidingclassificationoutputtoexecuting closed-loop automated UAV inspection, where the AI system dynamically adjusts flight paths and camera parameters based on real-time crack detection feedback withoutrequiringhumanoperatorinput.

Furthermore, the integration of advanced segmentationarchitecturessuchasU-NetandMaskR-CNN alongside the existing CNN framework presents a significant opportunity to expand from image-level classification to pixel-level crack delineation. This would enable precise quantification of crack width, area, and branching geometry, providing inputs for automated structural reliability calculations. Additionally, integration ofYOLOv8computervisionalgorithmsalongsidewearable IoT devices could automatically detect PPE violations, monitor worker fatigue, and identify dangerous proximity eventsnearheavyinspectionmachinery.

As data privacy and infrastructure security become increasingly critical, the development of cross-project federated learning models offers a highly promising research direction. This decentralised approach would allow AI models to securely learn from crack image data generated across multiple global inspection sites without compromising project-specific data regulations. Ultimately, expanding this framework to include BIM integration and Asset Management System connectivity will further automate maintenance scheduling, ensuring the continuous evolution of civil infrastructure inspection into a fully automated, resilient, and exceptionally safe discipline.

REFERENCES

1. Cha, Y.J., Choi, W., & Buyukozturk, O. (2017). "Deeplearning-based crack damage detection using convolutional neural networks." Computer-Aided Civil and Infrastructure Engineering,32(5),361–378.

2. Dhonde,H.B.,&Kalyanshetti,M.G.(2016)."Crack detection in concrete using image processing." Journal of Civil Engineering Research, 6(2), 45–52.

3. Mane, S., & Patil, V. (2019). "CNN-based pavement crack detection for Indian asphalt roads."IndianHighways,47(9),18–26.

4. Mohan, A., & Poobal, S. (2018). "Crack detection usingimageprocessing."AlexandriaEngineering Journal,57(2),787–798.

5. Padsumbiya, A., et al. (2022). "Lightweight CNN for crack detection on low-resolution images." Int.J.ofCivilEngineeringandTechnology,13(4), 22–31.

6. Rajesh, K., et al. (2024). "Enhanced YOLOv4-tiny for real-time crack detection in RC beams." Construction and Building Materials, 410, 134089.

7. Ramesh,N.,&Krishnan,S.(2018)."UAV-assisted bridge crack detection using machine learning." Journal of Bridge Engineering, 23(11), 04018080.

8. Shetty, R., et al. (2019). "CNN-based crack detection on Indian infrastructure images." Structures,18,211–220.

9. Sinha, S.K., & Fieguth, P.W. (2006). "Automated detection of cracks in buried concrete pipe images." Automation in Construction, 15(1), 58–72.

10. Zhang,L.,Yang,F.,Zhang,Y.D.,&Zhu,Y.J.(2016). "Road crack detection using deep convolutional neuralnetwork."IEEEICIP,3708–3712.

11. Bhelonde,A.,&Patil,S.(2020)."AutomatedCNNbased crack detection for Indian concrete infrastructures." Int. J. of Innovative Technology, 9(4),1234–1241.

12. Kumar, R., & Ghosh, A. (2022). "Dual-channel deep CNN for concrete crack detection." Construction and Building Materials, 321, 126352.

13. Bhosale, A., et al. (2022). "Pavement crack detection using deep learning on Indian roads." Journal of Transportation Engineering, 148(3), 04022003.