DCNN-BASED MODEL TO PREDICT CONCRETE STRENGTH FROM MOBILE PHONE IMAGES USING MACHINE LEARNING

Abstract

Concrete's compressive strength is critical for ensuring the safety, durability, and efficiency of construction projects. Traditional strength testing methods, though reliable, are labor-intensive, costly, and impractical for rapid, large-scale evaluations. This research explores a novel approach using Deep Convolutional Neural Networks (DCNN), specifically a ResNet50V2 architecture combined with transfer learning, to predict the compressive strength of concrete from images captured using mobile phones. A dataset of 49,000 dry concrete specimen images was prepared and enhanced through perceptual hashing for duplicate elimination, image preprocessing (resizing, normalization, standardization), and twenty-fold data augmentation, including random rotations, brightness/contrast adjustments, and horizontal flipping, to improve data diversity and model robustness. The DCNN model underwent a two-phase training process: initial feature extraction with frozen base layers, followed by fine-tuning of the final layers. The model accurately predicted compressive strengths of normal-strength concrete ranging from 12.96 MPa to 28.67 MPa. Evaluation metrics, including the coefficient of determination (R² = 0.9691), root mean square error (RMSE = 1.2349 MPa), and mean absolute percentage error (MAPE = 1.8651%), confirmed the model's high prediction accuracy. A paired t-test indicated no statistically significant difference (p = 0.9968) between true and predicted values, validating model reliability. An Android-based mobile application was developed for real-time, on-site predictions. A second paired t-test comparing outputs from the Python-based model and the mobile app yielded a p-value of 0.9148, confirming cross-platform consistency. This study presents a scalable, efficient, and highly accurate nondestructive method for concrete strength evaluation.