A Framework for Improving Object Recognition of Structural Components in Construction Site Photos Using Deep Learning Approaches

Abstract

Mobile phones and drones are widely used to record images of construction sites in various stages of progress. Although site images can be used a lot during the construction period, they are mostly used only for recording simple construction results. If objects in a photographic image are recognized by a component of construction structure, the image could serve as a valuable tool in construction schedule management. In this study, authors present a framework for automatically recognizing objects for each component of a structure in a photographic image. To this end, after collecting and training a large amount of photographic images from the railroad bridge structures, an object detection method that can automatically recognize the construction components of the bridge structure based on deep learning model was introduced. Images are procured through web crawling, and the collected images are pre-treated for supervised training. The results of the deep learning model showed high performance in the pier and coping classes, and the slab showed a rather low accuracy, and it was confirmed that the degree of utilization of the detection results was significantly affected by the angle of shooting the image.