Deep Learning Models for Crack Detection in RC Beams Using Thermal and Visual Imaging

This study investigates the effectiveness of deep learning models in detecting cracks in reinforced concrete (RC) beams using both thermal and visual imaging. A custom dataset comprising over ۲,۰۰۰ images, including high-resolution RGB and thermal infrared captures of RC beams under progressive loading, was developed under controlled laboratory conditions. Three convolutional neural network (CNN) architectures—ResNet۵۰, EfficientNetB۰, and a hybrid fusion model—were trained and evaluated to classify and localize crack patterns. Data preprocessing included thermal-to-RGB alignment, contrast enhancement, and image augmentation techniques to improve model robustness. The fusion-based model, integrating features from both modalities, outperformed single-input networks, achieving an average crack detection accuracy of ۹۴.۳%, with precision and recall values exceeding ۹۲%. Results highlight the complementary strengths of visual and thermal data, particularly in early crack detection under low-light or visually obstructed conditions. The proposed framework demonstrates strong potential for real-time, non-contact structural health monitoring of RC structures using AI-enhanced vision systems.

کلیدواژه ها:

Reinforced Concrete Beams ، Crack Detection ، Thermal Imaging ، Visual Imaging ، Deep Learning ، Convolutional Neural Networks ، Structural Health Monitoring