Concrete crack detection and severity assessment using deep learning and multispectral imagery analysis

Abstract

Traditional image-based crack detection methods often face limitations due to environmental noise, such as variable lighting, which impacts detection accuracy. This study presents an advanced approach that combines deep learning and spectral-index-based image analysis for concrete crack detection and severity assessment, addressing challenges in accuracy and stability. By employing the Single Shot Multibox Detector (SSD) trained on four types of spectral images—RGB, Global Environmental Monitoring Index (GEMI), Normalized Burn Ratio (NBR), and Normalized Difference Vegetation Index (NDVI)—the study demonstrates that detection accuracy significantly improves with multispectral imagery, especially for GEMI images, which achieved an average precision of 0.873. Regression analysis further reveals that crack orientation correlates more strongly with severity in transverse cracks than in longitudinal ones, providing critical information for automated maintenance strategies. A novel contribution of this work is integrating deep learning with multispectral data, representing a significant advancement in automated crack detection and offering enhanced detection precision and reliability in infrastructure health monitoring.