Visual Weld Quality Inspection System for Identifying Defects in GMAW Joints with Semantic Segmentation on Pre-trained Models

Gas metal arc welds are used in industrial manufacturing for joining metal components due to their high deposition rates, versatility, and cost-effectiveness. These processes are fundamental in sectors such as automotive, aerospace, construction, and notably, in the fabrication of tubular joints used for natural gas pipelines. Ensuring the integrity of these tubular welds is critical, as defects can lead to leaks, safety hazards, and costly failures in gas pipeline infrastructure, which may affect both human health and the environment. This research proposes an intelligent visual inspection system based on convolutional neural networks to detect and classify four critical welding defects (lack of continuity, excess penetration, spatter, and lack of penetration) through transfer learning with four pre-trained models (i.e., VGG-16, ResNet-101, EfficientNetB3, and Inception-ResNet-v2), integrated into three semantic segmentation architectures: U-Net; LinkNet, and Feature pyramid network. The methodology includes: (i) Preprocessing pipeline with image enhancement techniques applying contrast equalization technique enhancement to improve contrast and data augmentation techniques were employed to mitigate overfitting due to the limited dataset size; which is composed of 220 high-resolution images of stainless steel welding process created under controlled conditions, capturing both the top and root sides of the weld beads; (ii) Semantic segmentation was employed to perform pixel-wise classification of welding defects on visible-spectrum images of welding joints, enhancing the accuracy of automated visual inspection; and iii) Feature extraction was conducted using pre-trained convolutional backbones, capturing hierarchical visual patterns critical for identifying subtle weld defects, improving segmentation accuracy with reduced training data. The models were trained and evaluated across 96 configurations using accuracy, Jaccard index, F1-score, precision, sensitivity, and specificity as evaluation metrics. The results demonstrated that LinkNet, particularly when combined with the EfficientNetB3 backbone and the RMSprop optimizer, provided superior performance in segmenting exceed penetration defects, achieving an F1-score of 92.3% and specificity of 89.5%. For lack of penetration and lack of continuity, the U-Net and Feature pyramid network models showed competitive performance, with EfficientNetB3 again emerging as a robust and efficient backbone. These findings highlight the potential of deep learning-driven visual inspection for improving quality assurance in welding process applications, contributing to the development of smart manufacturing systems aligned with Industry 4.0 and sustainable development objectives.