Synthetic Defect Library

Organizes flaws into a hierarchical, machine-readable structure based on root cause, morphology, and severity. This moves beyond ad-hoc image collection to a structured ontology.

• Morphological Classes: Scratches, dents, cracks, inclusions, discoloration, and burrs.
• Root Cause Mapping: Links defects to upstream process failures like tool wear, thermal stress, or contamination.
• Severity Grading: Encodes dimensional tolerances and functional impact (e.g., cosmetic vs. structural).

Defects are not static images but procedurally generated assets defined by explicit parameters. This allows infinite variation from a single definition.

• Geometric Parameters: Length, depth, width, curvature, and fractal complexity of a scratch.
• Material Parameters: Bidirectional Reflectance Distribution Function (BRDF) properties to simulate how a defect interacts with light on metal, plastic, or glass.
• Stochastic Seeding: Reproducible randomization that generates diverse instances from the same defect class for robust edge case coverage.

Generates pixel-perfect labels simultaneously with the synthetic image, eliminating the cost and error of manual labeling. This is a critical advantage over real-world data collection.

• Segmentation Mask Generation: Provides exact pixel boundaries for defect regions.
• Bounding Box Synthesis: Creates tight 2D and 3D bounding boxes for object detection models.
• Depth Map Synthesis: Generates complementary depth data for training depth-aware inspection models.

Systematically varies environmental and sensor parameters to close the domain gap between synthetic training data and real factory-floor deployment.

• Camera Parameter Randomization: Varies focal length, sensor noise, lens distortion, and exposure.
• Environmental Variation: Randomizes lighting intensity, color temperature, and complex high-dynamic-range (HDR) backgrounds.
• Occlusion Modeling: Simulates realistic partial obstruction by tooling, fixtures, or other components.

Treats defect definitions as code, enabling strict versioning, reproducibility, and collaboration across engineering teams. This ensures model training is auditable.

• Semantic Versioning: Tracks changes to defect parameters and taxonomy over time.
• Provenance Tracking: Logs the exact generation parameters and seed values for every synthetic asset.
• CI/CD Integration: Allows automated validation of new defect types before merging into the master library.

Produces data across multiple sensor modalities from a single defect definition, enabling sensor fusion and robust perception.

• Photorealistic RGB: High-fidelity visual spectrum images using physics-based rendering.
• Infrared Thermal Maps: Simulates thermal signatures of subsurface defects like delamination.
• 3D Point Clouds: Generates structured-light or LiDAR representations for geometric inspection.

During NPI, physical defect samples are scarce or nonexistent. A synthetic defect library enables pre-production model training by injecting photorealistic flaws into pristine CAD renders.

• Zero-shot inspection on Day 1 of production
• Train on simulated scratches, dents, and flash before tooling is finalized
• Reduces NPI quality ramp time from weeks to days
• Validates camera placement and lighting angles in simulation

Catastrophic failures occur infrequently, leaving inspection models blind to their signatures. A synthetic library systematically generates long-tail anomalies that may appear once per million units.

• Simulate hairline fractures, inclusions, and delamination at scale
• Achieve 10,000+ labeled examples of defects that physically occur annually
• Prevent false-negative passes on safety-critical components
• Continuously expand the library as new failure modes are discovered in the field

Production environments introduce variability—lighting shifts, vibration, and contamination—that degrade model confidence. Synthetic libraries inject domain-randomized perturbations to force invariance.

• Vary lighting angle, intensity, and color temperature per defect instance
• Add simulated motion blur and sensor noise matching specific camera models
• Train against partial occlusion from coolant, chips, or adjacent parts
• Models maintain >99% recall under degraded real-world conditions

A centralized synthetic defect library provides a single source of truth for training across global factories, eliminating site-specific data bias.

• Version-controlled defect catalogs ensure reproducible training runs
• All sites train on identical failure mode taxonomies
• Enables federated fine-tuning where local real data augments the shared synthetic base
• Simplifies regulatory audit trails for inspection model qualification

Synthetic generation inherently produces pixel-perfect ground truth—segmentation masks, bounding boxes, and depth maps—at zero manual labeling cost.

• Every synthetic defect includes a precise segmentation mask for semantic segmentation training
• Bounding box coordinates are mathematically exact, not human-estimated
• Depth maps enable 3D defect characterization for volumetric flaws
• Eliminates the $5–$25 per-image cost of professional annotation services

A synthetic library serves as a regression test suite for inspection models. Before any model update deploys to production, it must pass against a held-out set of challenging synthetic defects.

• Define a golden validation set of edge-case anomalies
• Automate nightly evaluation runs in CI/CD pipelines
• Detect model drift or catastrophic forgetting before it reaches the factory floor
• Quantify recall per defect class with statistical confidence intervals