Welding Bead Detection and Inspection on Shear Reinforcement with Dual-Anchorage Using Enhanced Mask RCNN

Abstract

Welding processes pose significant challenges and risks to both humans and the environment, as well as the machines involved. To mitigate the potential negative impacts of welding, we suggest the development of a four-camera system that can replace human involvement in the assessment and evaluation of welded products. In this research paper, we introduce an algorithm based on Mask-RCNN and Green’s theorem-based classifier, designed to assess the quality of welding beads. For the optimization of our model, we trained the enhanced model on COCO weights using various backbone networks, including VGG16, ResNet101, MobileNet, Inception, and EfficientNet. Our findings revealed that Effi-cientNet reduced computational time compared to the other options. Due to constraints in data availability, we treated the problem as a single-class classification task. We addressed the issue of imbalanced datasets by incorporating a classifier with a decision threshold of 75% to distinguish between correctly and incorrectly welded beads. Furthermore , our proposed algorithm is versatile and adaptable, functioning effectively under different environmental conditions, with various types of cameras, diverse brightness settings, and multiple resolutions. Based on a thorough evaluation using performance metrics and the specified decision threshold, our algorithm achieved impressive results, a mean recall of 100%, precision of 100%, an accuracy of 99.56%, and an Intersection over Union (IoU) score of 98.92%. Additionally, the algorithm is capable of detecting and assessing the quality of welding beads within a mean processing time of 0.55 seconds per image, obtained from 172 frames. Based on these promising results, we recommend the adoption of our proposed algorithm for real-time applications that involve the detection and evaluation of welding bead quality, especially when integrated with the four-camera system we have developed. JEL Classification: D8 , H51