Massive-Scale construction dataset synthesis through Stable Diffusion for Machine learning training

Abstract

Advancements of artificial intelligence (AI)-driven image generation provide opportunities to address a problem in machine learning applications that have suffered from a lack of domain-specific training data. This study explores the feasibility of employing synthesized images (SIs) generated through Stable Diffusion as training data for construction. This study aims to examine the potential of Stable Diffusion in construction, and the performance of convolutional neural network (CNN) models trained exclusively on SIs. A total of 82.01% of images synthesized are suitable for representing construction tasks. The CNN model trained on preprocessed SIs (with context-based labeling results) exhibited a classification accuracy of 89.09%. The CNN model trained solely on raw SIs (synthesized images without context-based labeling results) achieved a successful classification rate of 86.51% for the images. This study presents the viability of SIs as a training dataset and introduces context-based labeling through object detection techniques, enhancing the performance of estimation models.