Automated Aging Building Defect Recognition and Analysis Using Transformers

Aging infrastructure poses significant safety risks due to the potential for undetected structural defects leading to catastrophic failures. Traditional manual inspection approaches are considered labor-intensive, time-consuming, and susceptible to human error. We propose a transformer-based defect detection framework specifically designed for detecting defects in aging buildings to solve the existing issues of manual methods. Transformer-based models rely on self-attention mechanisms to dynamically focus on relevant image regions. This approach avoids the limitations of conventional convolutional networks by capturing long-range contextual dependencies and adapting them to variations in defect appearance. To enhance the model’s generalizability and adaptability to real-world scenarios, a data preprocessing module is proposed to handle varying lighting conditions, angles, and environmental factors. Moreover, the proposed model integrates novel classification loss and matching cost to the state-of-the-art detection transformer (DETR) with the improved denoising anchor boxes (DINO) model to improve its performance and stabilize the learning process for accurate defect detection. The ability to analyze and visualize attention weights from multiscale features of the framework promotes transparency and builds confidence in automated defect detection systems. The experimental results reveal that the AD-TR framework outperforms competitive deep learning–-based object detection models with an average mean average precision (mAP) of 83.1%.