Pattern Recognition Using GAN-Based Feature Denoising

The patent applies local quality by introducing class-oriented dissimilarity loss that treats different classes differently in the feature transformation process. Instead of uniform error averaging, the loss function applies specific penalties for each class pair, making the transformation locally optimized for each class relationship. This resolves the contradiction by maintaining low within-class covariance while simultaneously optimizing class-oriented similarity through class-specific transformation criteria.

The patent changes parameters by introducing multiple loss components (GAN loss, classification loss, PLDA-based dissimilarity loss) that transform the feature space with different objectives. By adjusting the weights and combinations of these loss parameters, the system achieves both within-class compactness and between-class separation, resolving the contradiction through multi-parameter optimization rather than single-objective transformation.

The patent merges multiple transformation objectives into a unified framework by combining GAN-based noise robustness learning with classification-based separation learning and PLDA-based dissimilarity learning. This merging allows the feature transformation to simultaneously achieve noise robustness and optimal class separation, resolving the contradiction by integrating rather than prioritizing one objective over the other.

The patent implements feedback through the discriminator that provides classification feedback to the generator during training. The discriminator evaluates whether transformed features maintain class separability, and this feedback guides the generator to adjust transformations. This feedback mechanism ensures that noise robustness improvements do not compromise classification accuracy, as the transformation is continuously refined based on classification performance.

The patent segments the feature transformation objective into distinct loss components: GAN loss for noise modeling, classification loss for class separation, and PLDA-based dissimilarity loss for within-class compactness. Each segment addresses a specific aspect of feature quality, and their combination achieves high discriminative ability without excessive complexity, as each segment can be computed using established algorithms.

The patent creates a multi-functional transformation framework where the same transformation process simultaneously achieves noise removal, class separation, and within-class compactness. By making the transformation universal across multiple objectives rather than requiring separate transformations for each goal, the system maintains computational efficiency while achieving high feature discriminative ability through the synergistic effect of multiple loss functions.