Accurate aggregate classification, Particle Size Distribution (PSD) estimation, and concrete compressive strength prediction are critical for ensuring structural performance. This paper proposes a novel two-stage methodology for aggregate analysis and a new method for predicting concrete compressive strength, both leveraging DINOv2 embedding vectors extracted from aggregate images.
In our two-stage approach, we first use DINOv2 as a feature extractor to generate high-quality embedding vectors. These embeddings then serve as input for downstream classification and regression tasks (for aggregate type and PSD estimation) performed by small neural networks. This method achieves accuracy comparable to fine-tuned larger ViT-16 models dedicated for each task, highlighting the efficiency and potential of extracted embeddings. Furthermore, a Multi-Head architecture utilizing DINOv2 embeddings (both 384 and 768 dimensions) demonstrates improved performance over Single-Head MLPs. We also examine the impact of Patch Image Shuffling Augmentation on these models.
Extending these findings, we present an innovative strategy for concrete compressive strength prediction. This involves integrating the DINOv2 embedding vectors with data on concrete mix components. To address data scarcity, we introduce novel augmentation techniques for generating realistic mix designs. An ablation study provides compelling initial results, emphasizing the significant potential of this combined approach for predicting a range of concrete properties.
