.svg)
Automated welding lines require real-time inspection to catch defects like bad welds and porosity. The rarity of these defects makes it difficult to train reliable models, leading to costly false negatives. Issues like insufficient seat weld penetration in Tesla Model Y vehicles highlight the importance of rigorous inspection.
Role of Synthetic Data Solution
To address the shortage of rare weld defect examples, Advex AI created high-quality synthetic images showing cracks and porosity with different lighting and surface conditions.
Image: Advex AI automatically annotates with segmentation masks
Deep Dive 1: Bad Weld Defect
We augmented a model training set of 19 real images with 95 synthetic images. The results, shown in the chart below, revealed a fascinating insight into the roles of specialist and generalist models.
Image: Featuring Various Rare Weld Defects and Challenging Environmental Conditions
Performance and Insights: A Tale of Two Winners
The fine-tuned C-RADIOv2 foundation model delivered the best absolute performance, achieving the highest final mIoU score (0.72) with a 22.04% boost in IoU from synthetic data for welding defect segmentation. Next, we fine-tuned C-RADIOv2 with both the real + synthetic data for detection.
Graph: Bad Weld Detection Performance Uplift on C-RADIOv2
Importance of Model Selection:
The results reveals a key takeaway: C-RADIOv2 provided the best out-of-the-box result. However, the results also showcase the immense learning capacity of NVDINOv2 when paired with a SegFormer decoder, achieving the highest relative gain of 24.31%. For demonstrating the raw power of synthetic data to teach a model from a lower baseline, the generalist model told a more dramatic story with the larger dataset.
Graph: Bad Weld Detection mIoU Performance Uplift Across 3 Models
From Metrics to Reality: Visualized the Impact:
What does a mIoU score actually look like on the production line? It means moving from ambiguous outlines to clear, actionable detections, as illustrated below.
Image: Before and after comparison of a mask for a bad weld with a model enriched with Advex data
Deep Dive 2 : Porosity Defects
Among all weld defects, detecting surface porosity is particularly challenging. Its subtle nature and various potential causes make it easy to overlook.
The original data often had duplicate defects and lacked variation in porosity size, lighting, and weld textures. Advex AI's synthetic data specifically addresses this by generating a wide range of porosity examples.
Image: Advex Synthetic Data for Porosity Weld Defect
Performance and Insights: The Generalist Takes the Lead
For the subtle challenge of porosity, we augmented 32 real images with 160 synthetic examples. The results inverted the previous finding.
Winning Model Enhancement:
In a compelling reversal, the NVDINOv2 foundation model with the Segmentation head emerged as the clear champion. It not only achieved the highest performance gain with a 24.7% mIoU increase but also secured the top absolute score (0.5977).
Graph: NVDINOv2 model with Segformer Head improvement on Porosity Defects
Importance of Model Selection:
This outcome is a powerful lesson in the importance of empirical testing. The subtle and varied nature of porosity proved better suited to NVDINOv2.
Graph: Porosity Detection mIoU Performance Uplift Across 3 Models
From Metrics to Reality: Visualized the Impact:
What does a 24.7% increase in mIoU actually look like on the production line? It means moving from ambiguous outlines to clear, actionable detections, as illustrated below.
Image: Before/after comparison of a porosity detection mask
.png)
