Synthetic Defect Data Generation Using Latent Vector Mixing

The patent uses Generative Adversarial Networks (GANs) to generate synthetic defect images that copy the characteristics of real defects. The generator network creates artificial defect samples by learning from limited real defect data, effectively multiplying the training dataset without requiring additional physical defective products. This allows the machine learning model to be trained on sufficient diverse data while overcoming the low frequency of actual defect occurrence.

The patent transforms real defect images into augmented training samples by applying parameter changes such as rotation, scaling, cropping, and color adjustments. These transformations create variations of the same defect type, expanding the diversity and quantity of training data. The system modifies parameters like position, size, and orientation of defects to generate multiple training instances from limited original defect images.

Instead of collecting more real defect data through time-consuming inspection processes, the system copies existing defect characteristics through GAN-based synthetic image generation. The generator learns the distribution of real defects and produces artificial samples that preserve essential defect features while eliminating the need for extensive physical data collection. This dramatically reduces the time and resources required to build diverse training datasets.

The system performs preliminary action by pre-processing and augmenting available defect images before training begins. Defect images are pre-augmented with various transformations (rotation, scaling, translation) to create a expanded dataset in advance. This preliminary data preparation ensures sufficient training data availability without requiring extensive real-time data collection during the modeling process.

The patent transitions from fixed-rule algorithms to parameter-based machine learning models that can adapt to various defect types. Instead of hardcoding detection rules for specific defect shapes and sizes, the system uses learnable parameters in neural networks that automatically adjust to different defect characteristics. The model learns optimal detection parameters from training data, enabling flexible handling of diverse defect variations while maintaining implementation simplicity through automated training.

The system uses synthetic defect images generated by GANs to teach the machine learning model various defect patterns. By copying and transforming real defect characteristics into synthetic samples, the model learns to recognize diverse defect types without requiring complex hand-crafted rules. This approach maintains implementation simplicity while achieving high adaptability to varied defect shapes, sizes, and orientations.