CNN-CHD: Combining Clustering Hierarchical Divisive and CNN for Enhanced Weld Defect Detection

Abstract

Weld defects, such as cracks, porosity, and inclusions, can significantly compromise the structural integrity of welds, making their early and accurate detection crucial in various industrial sectors. In this research, we propose a comprehensive methodology that combines the Clustering Hierarchical Divisive (CHD) method with convolutional neural networks (CNNs) to enhance defect detection accuracy. Our approach begins with the creation of a robust database, leveraging Generative Adversarial Networks (GANs) for data augmentation, which allowed us to generate a more diverse and representative dataset essential for effective model training. The CHD method performs an initial segmentation of weld images, clustering them into coherent groups based on confusion matrix analysis, ensuring that each cluster corresponds to distinct defect classes. Subsequently, the clustered images are processed using CNNs, renowned for their powerful classification capabilities. This hybrid approach effectively captures the variability of weld defects, significantly improving detection accuracy while reducing similarities among defects. Our proposed CNN-CHD method offers a more efficient pipeline for defect identification in welding applications, and its potential to enhance quality control in industrial practices.