Performance and Efficiency Trade-offs in Image-Based Concrete Crack Detection: A Comparative Study of Traditional Machine Learning and Deep Learning Models

Resumen

Structural crack detection is a key task in structural health monitoring, as cracks often indicate early-stage damage in civil infrastructure. Traditional inspection methods rely on manual visual assessment, which is time-consuming, subjective, and potentially unsafe. To overcome these limitations, automated crack detection approaches based on machine learning and deep learning have gained increasing attention. This study presents a comparative analysis of traditional machine learning and deep learning models for image-based concrete crack detection. Handcrafted features extracted using the Histogram of Oriented Gradients (HOG) descriptor were classified using Support Vector Machine, Random Forest, and K-Nearest Neighbors, while deep learning approaches included a custom Convolutional Neural Network and a MobileNetV2-based transfer learning model. Experiments conducted on a balanced dataset of 40,000 images demonstrate that deep learning models achieve superior accuracy and robustness.