Comparative Study and Analysis of Deep Learning Models for Concrete Bridge Crack Detection

Abstract

Infrastructure for transportation relies heavily on concrete bridges, therefore maintaining their health is essential for everyone's safety. A comparison of deep-learning algorithms for spotting cracks in concrete bridges is presented in this work. The proposed models include Convolutional Neural Network (CNN), Convolutional Recurrent Neural Network (CRNN), You Only Look Once (YOLO) versions YOLOv5, YOLOv7, YOLOv8, and Residual Networks (ResNet) leverage cutting-edge deep learning architectures and feature engineering techniques, enabling more precise crack detection in concrete bridge structures. To boost model generalization and the capacity to spot cracks in a variety of real-world scenarios, various data augmentation techniques, such as Gaussian blur, mix-up, random rotation, center crop, random crop, Gaussian noise, random blocks, central region, and smart padding, were also included. The studies utilized cracked and uncracked concrete bridge surface photos from the open-source SDNET dataset. The accuracy, precision, recall, and F1 score of each model are evaluated. YOLOv8 has the highest accuracy of 95%, whereas CNN and YOLOv5 showed poor performance.