Interpretation and understanding of asphalt crack detection deep learning models using integrated gradient (I.G.) maps

Asphalt cracking poses a significant deterioration challenge in road networks. Historically, government agencies employed visual inspection methods to identify asphalt cracking. However, this approach is labour-intensive and time-consuming, prompting most authorities to transition to automated crack identification methods utilising Artificial Intelligence (A.I.). Convolutional neural networks (CNNs) have gained popularity due to their ease of implementation and enhanced accuracy. Nevertheless, CNN models were perceived as opaque entities. In this study, we employed Integrated Gradient (I.G.) maps to elucidate the workings of these models and interpret CNN-based crack image voxels that contributed to the positive (cracked) output of CNN. Our findings revealed that CNN utilised voxels within the cracks to classify the cracked image segments and highlighted the image features that CNN employed to distinguish cracks from their surroundings. These interpretation techniques provide insights into the behaviour of the CNN model, enabling the refinement of Deep Learning (DL) models.