Niveau : M.A.Sc.

A Computer Vision Approach for Quality Control and Defect Detection on Prefabricated Steel Elements


This research introduces an innovative computer vision approach for enhancing the process of quality control of welds in prefabricated steel elements, a crucial component in the construction industry. Traditional inspection methods, which are often expensive and time-consuming, are replaced by a proposed system that leverages cutting-edge computer vision algorithms and deep learning techniques. Our approach involves an instance segmentation model, co-developed with an industrial partner, trained on a diverse dataset of video frames. This model is designed to segment the weld area, measure its size in pixels, and isolate the shape for subsequent defect detection, thereby facilitating the identification of overflowed or underflowed welds. Despite certain limitations, our approach significantly transforms quality control practices, resulting in increased safety, reduced risks, and structural integrity. In our experiments, the instance segmentation model achieved an average precision of 94%. Furthermore, after testing the entire framework on real data and in a real environment, the proposed framework obtained an average accuracy of 90%. The results demonstrated the effectiveness of our proposed system in enhancing weld inspection processes.

Résultats du projet

The proposed instance segmentation network is a strong and reliable foundation for weld quality control and size-related detection. It accurately segments welds, isolates their shapes, and provides detailed information. By precisely identifying and delineating the weld areas, the network enables effective analysis and measurement of potential defects.
The results of the instance segmentation network were highly promising. It achieved an average precision of 94% on IoU 50, accurately segmenting weld areas from the background. To further validate the framework, the network was tested in a real factory setting. The entire framework underwent validation using two methods: shop validation and manual vs. automatic segmentation. On average, the accuracy of measurements reached an impressive 92%. These results demonstrate the reliability and accuracy of the framework in detecting and segmenting the weld areas, providing a robust solution for weld defect detection, size measurement, and quality control.

Contributions du projet

Research Contributions: 1) Developed a computer vision approach using novel deep learning techniques to automate and optimize the weld defect inspection process 2) Created a specialized instance segmentation model for accurate detection of weld areas, enabling the identification of potential defects and measurement of the weld size. 3) Prepared a diverse dataset of video frames depicting welded areas on construction joists for robust model training. 4) Identified limitations and proposed future research directions for practical real-time defect detection, exploring the utilization of novel technologies. Industrial Contributions: 1) Significantly reduced time and cost compared to conventional quality control methods, minimizing reliance on skilled labor and supervision. 2) Improved accuracy and processing time in weld defect detection, enhancing welding quality monitoring in the construction industry. This research combines academic rigor with practical industrial applications, enhancing both theoretical understanding and real-world impact in weld quality control.


L’équipe chargée de ce projet

Ali Motamedi


Ce projet a été supporté par :

Recherches similaires

Explorez plus en profondeur notre recherche en explorant ces études et ressources connexes :

Scroll to Top