Computer Vision For Defect Detection In Construction

Abstract

A new method is suggested here for detecting defects in construction by using PyTorch to implement a ViT-based semantic segmentation model. Problems such as cracks, corrosion and uneven surfaces on construction sites are challenging for people to check visually which is why new automated methods need to be used. Thanks to the self-attention mechanism, the ViT model is able to detect and locate faults on construction surfaces with great accuracy. Research shows that the new method delivers stronger results on average Intersection over Union (mIoU) and pixel accuracy compared to traditional CNNs and gives steady performance across diverse defects and conditions. Even though it requires careful training, the final model is fast at running and can be used right away at the work site. These results suggest that transformer-based networks could play a major role in developing quality control and monitoring applications in construction.