A Comparative Study of Deep Learning Models for Granulometry Image Based Estimation of Concrete Aggregate

Abstract: Obtaining the granulometry is the starting point of our pipeline for automating the calculation of concrete properties using images. For this reason, we focus on developing the best deep learning model that can compute aggregate gradation and generalize to images obtained from different aggregate producers. Therefore, we conduct a comparative analysis between different existing deep learning models and use three datasets to evaluate them : two publicly available and one of our own.Our analysis shows that transfer learning followed by fine-tuning on ViT_16 outperforms the other models, on both classification and regression tasks, with smaller errors and greater generalization capabilities.