Recent events in the automotive industry underscore the devastating impact of undetected cracks. When Honda had to recall high-pressure fuel pumps due to developing cracks, it highlighted a persistent challenge in manufacturing: traditional inspection systems often miss critical defects that appear similar to benign surface features. Manufacturers face a difficult choice - either wait months to collect sufficient defect data for AI training, or risk deploying unreliable models that could miss crucial flaws.

‍

Synthetic Data Solution

The original dataset contained only 8 real images with high variability in camera angles and surfaces. We used Advex to generate 40 synthetic images featuring diverse cracks, surface reflectivity, and lighting, creating a robust training set in just 40 minutes on NVIDIA H100s. Next, we fine-tuned C-RADIOv2 with both the real + synthetic data for detection.

‍

                                                                                                         _{Image: Advex-Generated Synthetic Images Featuring Variations in Crack Size and Location}

‍

Performance and Insights: A 73% Surge in Detection Accuracy with C-RADIOv2

‍

The results were remarkable. Using the specialist C-RADIOv2 model, we achieved a 73.41% improvement in detection accuracy, with the mIoU score jumping from 0.44 to 0.77. This demonstrates the immense impact of augmenting even a tiny dataset with high-quality synthetic examples.

‍

                                                                                                                                    _{Graph: C-RADIOv2 Performance Uplift on Crack inspection}

‍

Importance of Model Selection: 

This also highlights the importance of model testing and selection.  While SegFormer with NVDINOv2 and mit-b5 backbone also benefited from the additional synthetically generated training data, their gains were more modest compared to C-RADIOv2. A potential reason for this may be the fact that C-RADIOv2 is pre-trained in a supervised setting (vs NVDINOv2’s unsupervised pre-training) which enables more efficient learning from supervised fine-tuning.

‍

                                                                                                                      _{Graph: mIOU Score Comparison Across 3 TAO Models for Crack inspection}

‍

‍

From Metrics to Reality: Visualizing the Impact

A 73% increase in mIoU translates into a tangible difference in detection. The comparison below illustrates this principle: a model trained with synthetic data moves from hesitant, incomplete masks to higher confidence, precise masks.

‍

                                                                                                    ^{Image:  Before and after comparison of a mask for a crack detection with a model enriched with Advex data}

‍