Anomaly Detection Encoder Using Synthetic Image Data

The patent applies preliminary action by generating synthetic training data in advance through image processing techniques (such as adding noise, changing brightness, rotating, or flipping original images) before the actual training process. This allows the machine learning model to be trained with sufficient varied data without requiring extensive physical data collection, thereby reducing data collection time while maintaining training quality and anomaly detection accuracy.

The patent employs copying by creating synthetic copies of original images through various image processing transformations (noise addition, brightness adjustment, rotation, flipping). These processed copies serve as additional training data, allowing the model to learn from diverse variations without requiring proportional increases in physical data storage. This approach maintains model training quality while significantly reducing the volume of actual captured images that need to be stored and managed.

The patent applies parameter changes by systematically modifying image parameters (such as adding different types and levels of noise, adjusting brightness and contrast, applying rotation and flipping transformations) to original images. These parameter transformations generate diverse training samples that enable the machine learning model to learn robust anomaly detection capabilities even when the number of original captured images is limited, thereby maintaining detection capability while reducing data collection burden.