Abstract
Welding inspection plays an essential role in manufacturing industries to ensure the integrity and quality of weld joints. However, the prevalent manual inspection procedures are inherently subjective, prone to bias, and result in inconsistent quality assessments. Therefore, there is a strong need for an automated, intelligent system capable of objectively detecting welding spots. To address this, we propose an advanced segmentation model based on deep learning and computer vision techniques, specifically utilizing a Nested UNet (UNet++) architecture enhanced by extensive architectural modifications and comprehensive hyperparameter tuning. To further optimize segmentation performance, we systematically compare various convolutional blocks integrated into the bottleneck of the network architecture. Our experimental evaluation demonstrates that employing a VGG convolutional block at the bottleneck of Nested UNet achieves the highest performance, reaching an Intersection over Union (IoU) score of 76.18% and a validation loss of 0.1713 on our collected dataset. This research advances automated welding inspection by demonstrating that an enhanced Nested UNet with an optimized VGG bottleneck significantly improves weld spot segmentation accuracy. The proposed framework provides a robust benchmark for future studies on intelligent industrial visual inspection and deep learning-based quality assurance.