CNN-Based Crack Detection in Asphalt Pavements Using GPR: Intensity Map and Histogram Analysis

This study presents a proof of concept for using Convolutional Neural Networks (CNNs) to automatically detect cracks in asphalt pavements based on radargrams generated via Ground Penetrating Radar (GPR) simulations. All radargrams were synthetically produced through numerical modeling, varying crack widths (1 cm to 20 cm) and GPR frequencies (1.6 GHz and 2.3 GHz), with a fixed depth of 10 cm. Preprocessing included the application of a background removal smoothed filter to enhance structural visibility. Data augmentation techniques were applied to $\mathbf{1 0}$ original radargrams to address class imbalance, resulting in a dataset of 60 images (54 with cracks and 6 without). The CNN achieved an accuracy of 92.86 % on this balanced dataset. Intensity maps highlighted high-reflection areas in red and yellow corresponding to crack regions, while blue and green tones indicated homogeneous pavement. Histogram analysis revealed broader intensity ranges for cracked radargrams (typically 80-250) and narrower distributions for non-cracked ones (120-200), with some variations attributed to slab edge effects. Although limited to synthetic data, the results demonstrate the viability of combining GPR image processing and deep learning for crack detection. Future work should include experimental validation using real GPR data to assess the model's performance under practical conditions.