An innovative approach for aggregate gradation in old building mortars based on deep learning technique

Abstract

Aggregate gradation plays a crucial role in the mechanical and physical properties of mortars. In the restoration of old buildings, it is important to determine the aggregate gradation of mortars to maintain their original characteristics. However, traditional methods for analyzing the aggregate gradation of old mortars are both time-consuming and outdated. In this study, software was developed using state-of-the-art deep learning models for segmentation and object detection, allowing for faster, more accurate detection of aggregate distribution with less labor. This study introduces a novel approach using YOLOv8 deep learning models for automated, non-contact segmentation and detection of aggregates, voids, and cracks in mortar samples. Utilizing instance segmentation and object detection models from the Yolov8 family, the models achieved a score of 0.473 mAP mask50 for aggregate instance segmentation and 0.995 mAP mask50 for mortar segmentation on the test data, despite the small sample size. The models also delivered rapid results due to their short processing time and the development of a simple optical character recognition (OCR) method for scaling factor detection. This approach enables a faster and more efficient calculation of aggregate distribution, which can be applied to assess the quality of old concrete and analyze the characteristics of mortars used in historical building restoration. Fine-tuning with small, diverse datasets could enhance generalizability, broadening applicability to various structures. This approach advances intelligent construction technologies for efficient historical mortar analysis.