A Hybrid Image Processing and Clustering Framework for Unsupervised Crack Detection in Resistance Spot Welds

Resistance spot welding (RSW) inspection increasingly relies on supervised deep learning models; however, such approaches require large labeled datasets, extensive retraining under changing conditions, and often lack interpretability. Despite their high reported accuracy, there remains a clear need for robust, unsupervised, and training‐free crack detection methods that can operate reliably under varying surface textures and illumination conditions. To address this gap, this study presents a fully unsupervised computer vision pipeline for the automated detection and delineation of cracks in resistance spot welds. The proposed framework integrates median absolute deviation (MAD)‐based adaptive thresholding for statistically robust crack candidate extraction, Laplacian filtering for edge enhancement, and Density‐Based Spatial Clustering of Applications with Noise (DBSCAN) to isolate true crack regions from noise without prior training. A region‐growing strategy ensures crack continuity, followed by contour delineation for precise visualization. Experimental evaluation on representative weld images demonstrates that the pipeline effectively suppresses surface noise, isolates dominant crack structures, and produces continuous and accurately delineated crack boundaries without requiring labeled data. By eliminating dataset dependency while maintaining reliable detection performance, the proposed approach provides an interpretable, scalable, and cost‐effective alternative for automated nondestructive weld inspection, with strong potential for real‐time industrial deployment.