Pattern recognition apparatus, pattern recognition method, and computer-readable recording medium

A pattern recognition method for pattern recognition using neural network-based feature transformation, the pattern recognition method performed by a computer and comprising:

• generating denoised feature vectors by transforming noisy feature vectors into the denoised feature vectors;
• discriminating the denoised feature vectors and clean feature vectors corresponding to the denoised feature vectors by predicting probabilities that the denoised feature vectors and clean feature vectors corresponding to the denoised feature vectors are the clean feature vectors, and estimating corresponding classes of the denoised feature vectors and the clean feature vectors;
• calculating generator and discriminator losses using the denoised feature vectors, the clean feature vectors corresponding to the noisy feature vectors, the estimated corresponding classes of the denoised feature vectors and the clean feature vectors, and actual corresponding classes of the denoised feature vectors and the clean feature vectors, wherein
• the generator loss comprises a Generative Adversarial Network (GAN)-based loss, a classification loss, and a class-oriented dissimilarity loss,
• the discriminator loss comprises the GAN-based loss and the classification loss,
• the GAN-based loss is calculated using the predicted probabilities that the denoised feature vectors and the clean feature vectors are the clean feature vectors, the class-oriented dissimilarity loss is a dissimilarity between the denoised feature vectors and the clean features corresponding to the denoised feature vectors as measured using pre-trained probabilistic linear discriminant analysis (PLDA)-based-scoring, and
• the classification loss is a prediction error between the estimated corresponding classes of the denoised feature vectors and the clean feature vectors and the actual corresponding classes;
• and
• updating parameters used in generation of the denoised feature vectors and in discrimination of the denoised feature vectors and the clean feature vectors, such that the generator loss and the discriminator loss are minimized.

A non-transitory computer-readable medium storing a program for pattern recognition using neural network-based feature transformation, the program executable by a computer to perform:

• generating denoised feature vectors by transforming noisy feature vectors into the denoised feature vectors;
• discriminating the denoised feature vectors and clean feature vectors corresponding to the denoised feature vectors by predicting probabilities that the denoised feature vectors and clean feature vectors corresponding to the denoised feature vectors are the clean feature vectors, and estimating corresponding classes of the denoised feature vectors and the clean feature vectors;
• calculating generator and discriminator losses using the denoised feature vectors, the clean feature vectors corresponding to the noisy feature vectors, the estimated corresponding classes of the denoised feature vectors and the clean feature vectors, and actual corresponding classes of the denoised feature vectors and the clean feature vectors, wherein
• the generator loss comprises a Generative Adversarial Network (GAN)-based loss, a classification loss, and a class-oriented dissimilarity loss,
• the discriminator loss comprises the GAN-based loss and the classification loss,
• the GAN-based loss is calculated using the predicted probabilities that the denoised feature vectors and the clean feature vectors are the clean feature vectors, the class-oriented dissimilarity loss is a dissimilarity between the denoised feature vectors and the clean features corresponding to the denoised feature vectors as measured using pre-trained probabilistic linear discriminant analysis (PLDA)-based-scoring, and the classification loss is a prediction error between the estimated corresponding classes of the denoised feature vectors and the clean feature vectors and the actual corresponding classes;
• and
• updating parameters used in generation of the denoised feature vectors and in discrimination of the denoised feature vectors and the clean feature vectors, such that the generator loss and the discriminator loss are minimized.