Construction Quality Monitoring with Deep Learning-Based Automated Concrete Crack Detection: A Comparative Study

Abstract

Concrete cracking is a primary indicator of structural deterioration, yet traditional visual inspections are labor-intensive and subjective, limiting their effectiveness for large-scale monitoring. This study proposes an automated crack detection framework using Convolutional Neural Networks (CNNs) to address these challenges. Leveraging a balanced dataset of over 42,000 images, we conducted a systematic comparison of five transfer-learning models: VGG16, ResNet50, MobileNetV2, DenseNet121, and DenseNet201. Performance was evaluated across multiple metrics, including accuracy, F1-score, model size, and inference speed. The results identify DenseNet201 as the top performer, achieving a mean accuracy of 95.7% and an F1-score of 96.7%, driven by its superior stability across data partitions. Additionally, while MobileNetV2 offers a viable lightweight alternative for edge computing, DenseNet201 remains the most reliable architecture for high-precision quality monitoring tasks.