Automated Pavement Crack Detection Using Convolutional Neural Networks: A Comparative Study of Model Architectures and Preprocessing Techniques

Abstract

Pavement maintenance is essential for transportation infrastructure, ensuring road safety and efficiency. Traditional inspection methods are labor-intensive, time-consuming, and prone to human error. This research explores advanced image processing and deep learning techniques to automate the detection and segmentation of pavement cracks. The goal is to develop robust methodologies for accurately identifying and analyzing pavement cracks from high-resolution images, enhancing maintenance and repair strategies. High-resolution cameras mounted on vehicles and drones capture images of pavements under various conditions, including cracked and uncracked surfaces. These images are manually labelled and preprocessed using grayscale conversion, noise reduction, and histogram equalization to improve quality and prepare the data for model training. Data augmentation increases dataset size and variability. The study investigates different Convolutional Neural Network (CNN) architectures, including simple CNNs, enhanced CNNs with Batch Normalization and Dropout, and the pre-trained VGG16 model leveraging transfer learning. The simple CNN achieves a test accuracy of 85%, while the VGG16 model demonstrates superior feature extraction and crack detection performance. Hyperparameter tuning, particularly the learning rate, is critical, with an optimal rate identified at 0.0035 for enhanced CNNs. Evaluation metrics like accuracy, precision, recall, F1-score, and Intersection over Union (IoU) offer a comprehensive assessment of model performance.