Defect Detection Model Using Normalized Loss Functions

The patent applies preliminary action by performing data augmentation and normalization preprocessing on normal samples before training the generative adversarial network. This includes adjusting the distribution of training data, normalizing color spaces, and pre-processing images to ensure consistent input quality, which improves detection accuracy while maintaining the semi-supervised approach that requires minimal anomaly samples.

The patent employs parameter changes by modifying the loss function parameters and training hyperparameters of the generative adversarial network. Specifically, it adjusts the weighting parameters in the composite loss function (combining reconstruction loss, adversarial loss, and normalization loss) and optimizes learning rates and batch sizes to balance training stability with detection precision, thereby resolving the accuracy issue while keeping data collection requirements low.

The patent applies segmentation by dividing the training process into multiple stages with different objectives and hyperparameters. The first stage focuses on basic reconstruction capability, the second stage introduces adversarial training, and the third stage applies normalization constraints. This staged approach accelerates convergence and improves stability compared to single-stage training, while the modular structure keeps the overall complexity manageable.

The patent implements dynamics by making the training process adaptive through dynamic hyperparameter adjustment across stages. Learning rates, loss function weights, and batch sizes are dynamically modified based on training progress and performance metrics. This dynamic adjustment optimizes training speed and stability without requiring overly complex fixed architectures.