Data Augmentation

Apply spatial modifications to defect images to simulate varying part orientation and camera positioning on the production line.

• Rotation: Randomly rotates images within a defined range (e.g., ±15°) to handle misaligned components on conveyor belts.
• Scaling & Zooming: Simulates parts at different distances from the camera, teaching scale invariance.
• Flipping: Horizontal and vertical mirroring, critical for symmetrical parts where defect orientation is irrelevant.
• Shearing & Elastic Deformation: Distorts images to mimic flexible materials or slight camera lens distortion.

Example: A scratch defect on a metal casing must be detected whether the part enters the inspection station rotated 0° or 45°.

Alter pixel intensity values to replicate the unpredictable lighting conditions common in industrial environments.

• Brightness & Contrast Adjustment: Simulates fluctuating ambient light, aging LED illuminators, or shadows from nearby machinery.
• Hue & Saturation Shifts: Accounts for variations in part surface finish, oxidation, or inconsistent paint batches.
• Gaussian Noise Injection: Mimics sensor noise from high-speed line scan cameras operating at low exposure times.
• Channel Shuffling: For multi-spectral imaging, randomly dropping or swapping color channels forces reliance on shape over color.

Goal: Prevent the model from using absolute pixel intensity as a shortcut, ensuring it focuses on the structural features of a defect.

Simulate partial visibility and overlapping objects to prevent the model from relying on a single, unobstructed view of a defect.

• Random Erasing (Cutout): Removes a random rectangular region of the image, forcing the network to learn from surrounding context. This mimics a defect partially covered by debris or coolant.
• Mixup: Blends two training images and their labels linearly, creating a virtual example that lies between classes. This regularizes the decision boundary against adversarial noise.
• Mosaic Augmentation: Stitches four training images into one, a technique popularized by YOLOv4 to improve detection of small defects in complex scenes.

Benefit: Drastically reduces the False Reject Rate (FRR) caused by transient obstructions like water droplets or dust on the lens.

Use generative models to create entirely new, photorealistic defect images, bypassing the physical impossibility of collecting thousands of rare failure samples.

• Generative Adversarial Networks (GANs): A generator network creates fake defect images while a discriminator learns to distinguish them from real ones. The resulting generator produces high-fidelity anomalies like cracks, dents, or contamination.
• Diffusion Models: State-of-the-art models that iteratively denoise random patterns into controlled, high-resolution synthetic defects with precise textural detail.
• Procedural Generation: Algorithmically renders defects using physics-based rules (e.g., simulating a splatter pattern or a scratch path) for deterministic control over defect morphology.

Application: Training a model to detect a rare "edge chip" defect that occurs in 0.001% of units, a scenario impossible to capture naturally at scale.

Apply augmentation not just during training, but also during inference to produce a more stable and accurate prediction for a single query image.

• Multi-View Consensus: The input image is augmented into multiple variants (e.g., original, flipped, rotated). The model runs inference on all variants, and the final prediction is an average or majority vote of the outputs.
• Stabilization Effect: TTA reduces prediction variance caused by ambiguous edge cases, effectively acting as an ensemble without training multiple models.
• Computational Trade-off: Increases inference latency by a factor equal to the number of augmentations applied, making it suitable for offline batch inspection or when deployed on high-throughput Neural Processing Units (NPUs).

Result: A measurable improvement in Mean Average Precision (mAP) without any changes to the model architecture or retraining.

Systematically vary the entire visual environment of a scene—including backgrounds, lighting positions, and textures—to bridge the gap between synthetic training data and the chaotic reality of a factory floor.

• Background Swapping: Pastes defect objects onto hundreds of random background textures (metal, plastic, wood grain) to prevent overfitting to a specific conveyor belt color.
• Lighting Randomization: In a simulated environment, randomizes the number, position, color temperature, and intensity of virtual light sources.
• Sim-to-Real Transfer: A foundational technique for training models entirely in simulation (e.g., NVIDIA Omniverse) and deploying them directly to physical hardware without real-world fine-tuning.

Key Metric: Minimizes the Escape Rate by ensuring the model has been exposed to extreme visual diversity before encountering a real, non-conforming product.