Transfer Learning-Based Object Detection Model for Steel Structure Bolt Fastening Inspection

Abstract

As improper inspection of construction works can cause an increase in project costs and a decrease in project quality, construction inspection is considered a critical factor for project success. While traditional inspection tasks are still mainly labor-intensive and time-consuming, computer vision has the potential to revolutionize the construction inspection process by providing more efficient and effective ways to monitor the progress and quality of construction projects. However, previous studies have also indicated that the performance of vision-based site monitoring heavily relies on the volume of training data. To address the issues of challenging data collection at construction sites, this study developed models using transfer learning-based object detection models incorporating data augmentation and transfer learning. The performance of three object detection algorithms was compared based on average precision and inference time for detecting T/S bolt fastening of steel structure. Despite the limited training data available, the model’s performance was improved through data augmentation and transfer learning. The proposed inspection model can increase the efficiency of quality control works for building construction projects and the safety of inspectors.