Assessing the Viability of Generative AI-Created Construction Scaffolding for Deep Learning-Based Image Segmentation

Construction scaffolding serves as a pivotal temporary structure essential for construction activities, exerting a direct influence on site safety conditions. Unfortunately, the lack of documentation often leads to a shortage of training data necessary for employing image segmentation through deep learning for inspection purposes. In an effort to overcome this bottleneck, Generative AI, adept at creating images from pretrained data, emerges as a potential solution. However, the inherent black box nature of deep learning introduces the possibility of generating unrealistic images, thereb necessitating a rigorous evaluation, which constitutes the primary focus of our research. Our findings reveal that scaffolding images generated by Generative AI exhibit distinct features that our deep learning model successfully learned, resulting in an impressive mean average precision (mAP) of 82. Nonetheless, discernible patterns in image generation may be lacking, as evidenced by our deep learning system's ability to grasp scaffolding features proficiently, achieving a mAP of 69 even from the initial epoch. This observation suggests potential challenges in generating diverse scaffolding images through the Generative AI approach, emphasizing the need for further investigation before implementing it with real scenario images