| Title |
A Study on Construction Material Recognition Using YOLO and Virtual Images |
| Authors |
Hyunwoo Kim ; Jeongseop Kim ; Minkoo Kim |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2025.26.5.090 |
| Keywords |
Construction Material; YOLOv8; DALL-E; Virtual Images; Object Detection |
| Abstract |
This study presents a construction material recognition method using virtual images to enhance safety and productivity at construction sites. Real and virtual images for eight major materials were collected using Pythonbased web crawling and OpenAI’s DALL-E, with data augmentation techniques applied to construct training data. Using the YOLOv8 object recognition algorithm, the study analyzed model performance based on the proportion of virtual images used. The results are as follows (i) The model using only real images had an mAP@0.5 of 0.765, while incorporating approximately 75% virtual images increased the mAP@0.5 to 0.830?an 8.5% improvement. This highlights the importance of appropriately combining virtual and real images to enhance model performance. (ii) Recognition rates for materials like Cement Bag, Wood Plank, Glass, and Stone significantly improved with the inclusion of virtual images, compensating for the lack of diversity in real images. Notably, the mAP@0.5 for Glass increased from 0.368 to 0.740 (101.09% improvement), and for Cement Bag from 0.547 to 0.721 (31.81% improvement). (iii) For materials such as Brick, Rebar, Sand, and Pipe, adding virtual images had little effects or slightly decreased performance, indicating that virtual images are more beneficial for materials with limited data or diversity. The study confirms the potential of using virtual images for construction material recognition, contributing to the advancement of smart construction technology. |