This research addresses the critical aspect of accurately identifying weld defects, pivotal for weld quality and economical production. It introduces an innovative approach that synergizes image features with machine learning algorithms to discern diverse weld defects. The investigation evaluates the performance of three distinct feature extraction methods, encompassing the Histogram of Gradients, the Local Binary Pattern, and the Pooling layer from the ResNet18 Neural Network. These feature extraction techniques are combined with classifiers, including support vector machines, K-nearest neighbors, and decision trees. An external dataset is employed to validate the proposed model, encompassing five types of weld defects. The experimental outcomes demonstrate the superiority of the pre-trained ResNet18’s pooling layer feature extractor when coupled with a support vector machine classifier. This configuration yields an impressive classification accuracy of 99.75% via tenfold cross-validation and processes images at 3200 observations per second. This study's findings underscore the efficacy of the presented methodology in accurately identifying weld defects and their types, consequently contributing to improved weld quality and operational efficiency within the realm of intelligent robotic welding applications.